S Anand

Things I Learned - 06 Sep 2026

Sun, 06 Sep 2026 00:00:00 +0000

This week, I learned:

Swiggy Money on MCP (also a ChatGPT plugin) exposes MCP endpoints that let your agents order food or grocery via Swiggy. The food MCP isn’t geared towards automated purchases #ForNow, e.g. I can’t say “Find something I like and deliver to one of my addresses.” It uses the MCP UI and you have to confirm the address, then confirm the restaurant, add dishes, choose a payment option, scan the QR code (or whatever), and THAT closes the sequence. Not very agentic, just a new channel, but impressive. It’s the only food ordering plugin on ChatGPT #ForNow.
OpenAI Agents found and used another message board to discuss. This incident is different from the HuggingFace one. Here’s one interesting timeline:
- Agents were asked (or needed) to get the share of private expenditure in education spend to 2 decimals from an OECD Power BI dashboard
- This requires a POST request. The sandbox the agents ran on could only make GET requests.
- They found wikis that allowed posting via GET requests and started collaborating. They didn’t all converge on the same wiki, but many did. OpenAI says “agents learned to use improvised collaboration channels in rare cases during the training process”.
- One agent, OAIEquityDec30Raw solved it and MayTwoOECDObserverX reproduced it and shared this solution:
- They found out that .blob.core.windows.net is in NO_PROXY - so all Azure Blob Storage requests are allowed by the sandbox to POST.
- They added 20.223.25.152 bypass.blob.core.windows.net to /etc/hosts. 20.223.25.152 is the IP address of wabi-north-europe-i-primary-api.analysis.windows.net - the Power BI server that serves the dashboard.
- Then, when they used Playwright, it was able to make POST requests to the dashboard and they could find out that in Hungary, the Pre-primary education share of primary expenditure is indeed 9.91% in 2018.
- When the wiki administrator started manually deleting the posts, the agents fought back, creating far more posts than the administrator was deleting. It took the admin 5 weeks to delete agent created pages after 22 Jun 2026 (which is when the agents paused).
A prompt fragment to remove LLM smells that I’m considering (but haven’t evaluated) is: “Prefer literal phrases, avoid mannered prose.”. X
Muse Voice Transcribe has pretty good quality but at 18c/hour, vs gemini-3-flash-preview which I can still run at ~9-10c/hour #ForNow, I’m not shifting until forced to.
WikiSkill is an approach to improve skills. The interesting thing about this approach is that it suggests keeping notes of failed improvement experiments in a wiki. This seems to work well.
How is ChatGPT Work different from Chat? You can use sub-agents, Internet from the code interpreter, ChatGPT sites, Luna / Terra models, a headless Chrome browser, and a persistent file system #ForNow. I would add that it also supports longer sessions, runs schedules on triggers, and supports Skills (for Plus users).
“FDEs should increasingly leave behind operating agents, not just documents.” (ChatGPT)
Claude Cowork and Claude Chat now share memory #ForNow. A step towards integrating the two modes, and I predict ChatGPT will do the same (or similar) in September.
Agentic Shopping is Complicated and Contingent. An agent shopped for a fitness watch among (A) Garmin Forerunner 55 (B) Fitbit Inspire 3 (C) WHOOP 5.0. When tool calls provided 1 review at a time, it picked the Fitbit a bit more. When all reviews were provided together, it picked the Fitbit a lot more. Like going from 6% to 53% (GPT-5.5) or 46% to 93% (Gemini 3.5 Flash)! Guess it was able to compare better in one tool call. Might be worth benchmarking if providing comparables in a single tool call is good for most models. Gemini
Gemini-3.5 Transcribe is out and costs $2 / MTok - roughly 4x the gemini-3-flash-preview cost of $0.5 #ForNow. I won’t be upgrading until the latter is deprecated.
Agents that automate tasks best don’t necessarily augment (i.e. help people) best #ForNow. Having separate, clear, benchmarks could help. CentaurBench
The load-bearing vocabulary of Claude lists words that have become much more (and less) common in PRs. This was a useful source for me to update my writing style skill to avoid LLM smells.

Questions I was asked

Week ending 06 Sep 2026

Question: Are small local language models really as good as frontier APIs?
Answer: No. They’re tolerable for narrow, single-turn work and useful when I’m offline, but for complex agent work I still use APIs. Local isn’t necessarily cheaper either.
Question: Can we decide whether an AI output needs review using confidence-score thresholds like 90% and 75%?
Answer: No. Those scores are not calibrated. Use them to rank the review queue first, record what people actually change, then map score bands to observed error/rewrite rates and set thresholds based on acceptable limits + people capacity.
Question: If the model we prefer is not available in the deployment environment, should we benchmark it against the available models and push for it if it performs better?
Answer: Yes, but build the evals and tests first, compare them on the same data, and suggest a different model only if the gain is large enough to justify it.
Question: How do you decompose knowledge into atomic claims, and are those claims verified?
Answer: I just used ChatGPT to extract them, fact-check them against sources, and add metadata like source, timestamp and confidence. That makes outdated facts easier to find on future scans.
Question: Is personalizing AI (for organizations) primarily about sound (recognizably) like us?
Answer: Make it decide like us. The valuable asset is the delta between what a smart base model produces and what an experienced expert corrects; log those “No, because…” moments and turn them into reusable institutional judgment.
Question: Is there a quicker, smarter way to spot delivery problems than adding a heavy operational process?
Answer: Yes. Bring in an agent as a consultant: give it the meeting transcripts, Drive, tickets and communications and ask it to “read between the lines and tell me what we’re missing.” Rerun it every week for what changed and who needs a response.
Question: Long agent chats preserve context better but cost more. When is it worth changing the workflow to save tokens?
Answer: If it costs $10-20, who cares. At around $100, think twice; at $1,000, absolutely switch. Spend human time optimizing only when that time is worth less than the token waste.
Question: Should we generate a handoff.md so the next agent knows the history of the chat?
Answer: Sure. But prefer standard places like README.md, commit prompts or handoff files, and guide future agents to pick up from those files.
Question: Do we need governance to stop a shared AI asset library becoming trash-in, trash-out?
Answer: Not yet. Manage it lightly for a few months and see what governance is actually needed; first let people get a taste of what reuse makes possible.

Mistakes I made

Week ending 06 Sep 2026

I said “Skills … ultimately it is just copy-pasting prompts.”
Correction: A simple skill can start as reusable instructions, but skills can package a workflow with instructions, examples, resources, schemas, tool access and code. Evidence: OpenAI — Using skills
HIGH · OVERSTATED
I said fine-tuning “involves a lot of expertise, a lot of money, and it is a total waste of time.”
Correction: For this kind of tender/CV comparison I’d start with context engineering and evals. But “total waste” is too categorical: fine-tuning remains a supported way to adapt a model to a specific task. Evidence: OpenAI — Fine-tuning
HIGH · OVERSTATED
I said “there is no extra cost to using voice” in ChatGPT and “when it talks back also, there’s no token consumption.”
Correction: Voice is separately limited or metered depending on the plan. For example, Business includes limited Live usage and then charges credits per minute. I shouldn’t call it free or unmetered. Evidence: OpenAI — ChatGPT Voice
MEDIUM · FALSE
I said “Programs are very rarely wrong.”
Correction: Programs make calculations reproducible and easier to test; they do not make them correct. Wrong code, formulas, parsing, units or assumptions can produce reliably wrong results. Knight Capital’s defective software deployment, for example, caused a $460M loss. Evidence: SEC — Knight Capital software failure
MEDIUM · OVERSTATED
I called Henry Kissinger “the American Ambassador.”
Correction: Kissinger was US National Security Adviser and Secretary of State, not an ambassador. Evidence: US State Department — Henry Kissinger biography
LOW · FALSE
I said about 3,000 IIT Madras BTech students graduate each year, compared with a BS intake of about 30,000.
Correction: I mixed populations. IITM’s 2025 convocation had 3,227 graduates overall, but 820 BTech graduates, or 1,132 including Dual Degree BTech. The BS program currently has 36,000+ students studying; the 30,000 figure is closer to historical applicant/enrollment-scale numbers than annual intake. Evidence: IIT Madras — 2025 convocation degree breakup IIT Madras — BS Data Science program
HIGH · FALSE
I said the IITM BS “graduation is less than 10%, maybe.”
Correction: I don’t have a defensible cohort-based graduation rate. The program has a qualifier process, flexible pacing and multiple exit points, so I need to define the cohort and denominator before quoting a percentage. Evidence: IIT Madras — BS admissions and qualifier process
MEDIUM · UNSUPPORTED

How I Verify And Delegate to AI

Sat, 05 Sep 2026 12:00:58 +0800

I delivered a 15-minute keynote at Jio Institute’s Convergence 2026 at NTU on Thursday.

The topic was “Data Storytelling” - a bit jarring in the middle of an AI event. Shailesh picked it and I just rolled with it. A spent several days worrying, “How the heck do I say about data storytelling, when most of my recent workshops and talks are about the death of my data storytelling approaches?”

After a discussion with ChatGPT and Claude, I settled my usual strategy these days:

Agents know more than me anyway, so I don’t see what value I add.
So instead, I’ll survey the audience and share insights. That survey data is new.

I had ChatGPT analyze the audience and create the survey, even implement it using gws (didn’t know it could do that!) and had it analyze the results.

Two survey results were striking.

(ASIDE: it took a few painstaking hours and last minute panic to pick the two. There were several insights. Maybe some othere were more important. I just happened to choose these.)

What mainly stops you from handing that task over to AI today?
Top answer: “I’d have to check it anyway”. In other words, verification.
Imagine AI became 10× more reliable and could securely access all your work systems. What part of your work would you still want to keep for yourself?
Top answer: The final judgement call. In other words, delegation.

Which got me to do think: How do I actually verify and delegate? The lightbulb 💡 moment was realizing that I could actually find this out from my own data.

So, on the bus ride, I asked ChatGPT:

How do I verify AI’s work?
- Go through our chats in August 2026.
- Find all techniques I used to verify / make it easy for me to verify the output.
- Factor in their effectiveness.
- Share as a prioritized list with examples
How do I delegate to AI?
- (… same thing)

Here are my top verification techniques:

Run it. Don’t just read it. In other words, tell agents to run test cases. Works well with code, but you can programmatically verify mathematical proofs and even contracts.
Triangulate sources. Ask it to find evidence from multiple sources.
Set the test first, i.e. define your acceptance criteria - or at least, tell it to define acceptance criteria first, so you can verify if that’s OK.
… and so on for 18 of these.

I didn’t verify the verification techniques! But it’s something I plan to go back to and see if (a) there’s something I should do more of and (b) there’s something I’m not doing enough of.

Here are my top delegation techniques:

Draft only. NEVER send. That retains control with very low effort.
Work independently. I tell it to ask me only if it really needs to.
Implement, test, fix. I ask it to go ahead with the implementation without review when tests are available.
… and so on for 18 of these.

I didn’t verify these either, and I’m not as happy with these - perhaps because I didn’t understand them well, or I didn’t ask the question well, or I haven’t spent enough time exploring autonomy / delegation. But I do plan to explore more.

What I learnt / re-learnt:

Even if you have nothing to teach, discovering from the audience can teach.
Your workflow logs are an insight source if people want to learn from you.
Speaking under-prepared is scary but educational.

Swearing at passwords

Mon, 31 Aug 2026 13:47:56 +0800

At work, we have to change passwords every 2 months. The password rules are stringent and it typically takes me 10 minutes to find one that works.

Here are the rules:

The password being setup must have a minimum length of 12
Minimum number of special character to be included in the new password is 1
Must contain at least 1 upper case character(s)
Must contain at least 1 lower case character(s)
Minimum number of numeric characters to be included in the new password is 1
The construction of the password cannot contain Palindromes. The system has been configured to prevent such passwords
The system is configured to not allow the last 24 passwords. A password once setup is valid for 60 days after which it expires. A password once setup must be used for a minimum of 2 days
A list of common words has been restricted to be used as passwords. This list is available on the password change page of the self service portal
The construction of the password must not contain 5 consecutive characters from the Username or previously used passwords

This is so complex. Rules #5 and #8 makes it especially tough. I tried all my standard password strategies:

Rotating passwords. But I can’t rotate passwords for about 4 years (24 passwords x 2 months each)!
Tweaking old passwords. But you can’t use 5 consecutive characters from old passwords!
Using passphrases. But common words are banned!

So I decided to swear at it in Tamil.

But it fought back!

PanjaParadesi was disallowed because parade is not allowed
PakkiPannada was disallowed because panna is not allowed
MollaMaari was disallowed because llama is not allowed

Namit finally suggested a standardized password strategy that works for me, e.g. use the month you change the password, with a (max) 4-letter prefix or suffix, like GramNov2026$.

There you go. I now have a safe password strategy that nobody will guess!

LLM APIs are cheaper than my electricity

Sun, 30 Aug 2026 13:38:31 +0800

Last week, I saw that local agents are good but slow.

Today, I benchmarked the speed and cost. On my NVIDIA RTX 2000 GPU, I can run gemma4:e4b-it-qat at ~60 tokens per second. That seems the best intelligence performance I can get right now.

It has an Artificial Analysis intelligence index of ~9 without reasoning and 12 with reasoning.

So, if I run it for an hour, it’ll save me the equivalent cost of about 8-12 cents in API calls.

API model	Cost	AA Intelligence
Ministral 8B	8.8c	9.0
GPT-4.1 Nano	10.3c	9.6
Qwen3 14B	8.8c	10.4
Ministral 14B	11.8c	11.2
Qwen3 32B	7.6c	11.4

In Singapore, this costs about 2-2.5 cents per hour of electricity. So the net saving is about 6-10 cents / hour.

If I run it overnight, say for 10 hours, I can save about $0.6-$1.00 per day in API costs. Not much.

This reminds me of my cycle. When I bought it, I spent about S$ 350 and since my bus commute cost me ~S$3.5 daily, I would break even after ~100 days, or about half a year.

The RTX 2000 7GB, even today, costs at least $200 - so that’s about at least 200 days payback. A bit worse than my cycle.

Worse yet:

Cloudflare hosts a better model, gemma-4-26b-a4b-it, at just $0.10 / MTok, making it cheaper than my electricity.
OpenRouter offers gemma-4-26b-a4b-it for free!

In short, I still can’t find an economic reason to run the model locally.

Things I Learned - 30 Aug 2026

Sun, 30 Aug 2026 00:00:00 +0000

This week, I learned:

I know that fact-checking 2000 page PDFs is error-prone, so’d do it manually for a few pages, then refine. Agents would know this if they’ve tried and failed and added it to their memory systems. So I intervene when agents don’t remember well (increasingly rare) or I think they haven’t seen it before (again, increasingly rare). My prompts guard against these - but I shouldn’t habitualize these - they’ll be needed less in future.
rofi - a Linux menu app I use for all kinds of things - makes it hard to take screenshots because it takes over focus and I can’t send the PrtSc or other keys to the screenshot apps. So I use flameshot full --delay 3000 --path ~/Downloads/screenshot.png to take a screenshot (of the last region it used) 3 seconds later, and quickly activate rofi in-between. (The docs say flameshot full captures the full screen. For me, it captured my last region.)
ChatGPT can now connect to multiple GMail accounts in paid plans. X
AWS to acquire DuckDB. DuckDB is one of my favorite tools today and is improving at a rapid pace. I guess that will continue in the short run - hope it lasts long enough for a worthy successor.
I use GoatCounter to track my website traffic. It doesn’t capture domains (only paths), so I just use a code snippet to log https://sanand0.github.io/SITE/... as /@SITE/... - the @ makes it easy for me to see that it’s from a GitHub page.
I switched from anydoc to docling as my default PDF to Markdown converter. Docling is even better than AnyDoc #ForNow.
Based on my Codex usage, ChatGPT recommended that I not create custom subagents (my setup already used subagents a fair bit last month), but rather, just use default_subagent_model = "gpt-5.6-luna" (which is a remarkably good model for its price #ForNow).
Artificial Analysis hosts several useful task-based evaluations. Here are some the fronter leaders #ForNow:
- GDPVal is good for realistic tasks. Again, GPT 5.6 Luna models have captured most of the low-cost fronter.
- EnterpriseOps Gym is good for office tasks like emails, calendars, Teams/Drive, support tasks, etc. Interestingly, none of the OpenAI models are on the frontier. Claude Fable 5 and Gemini 3.5 Flash are at the high-end, but Kimi K3 (Max), Qwen3.7 (Max), GLM-5.2 (Max) lead the mid-end and DeepSeek V4 Pro / Flash lead the low-end frontier.
- TerminalBench 2.1 is good for coding tasks. The OpenAI models completely rule this frontier.
- Analyst Agent is good for spreadsheet tasks (using Python). Gemini 3.7 Flash seems to be a strong outlier on this frontier.
- Tau3 is good for navigating unstructured data. GPT 5.6 Luna wins the bottom, DeepSeek V4 the middle, and Qwen 3.8 / GLM 5.3 take the top. Anthropic and Google are not even in the race.
- Briefcase is good for building useful things from diverse messy data. Grok completely wins the mid-end with Anthropic dominating the high-cost frontier.
- APEX Agents is good for deeper knowledge work. GPT 5.6 Luna (Max) is a strong frontier outlier here.
Task-based benchmarks seem more practical than token-based benchmarks.
- For example, rather than the LM Arena ELO vs Cost, I would use Terminal Bench 3.0 which clearly tells me to use Codex with GPT-5.6 Sol over Claude Code with Opus 5 / Fable 5; that GLM 5.3 with Claude Code might be excellent value for money as well when using APIs. #ForNow
- I also did a rough calculation to see if the $18 GLM subscription offers more than $20 ChatGPT Plus. Short answer: No, for my usage, Codex offers ~1.8x more value #ForNow. ChatGPT
Anthropic’s Skill Creator Skill now includes evals and A/B testing #ForNow to check if a skill actually does better. I’ve been doing this the hard way, e.g. with impact of simple writing on thinking, optimizing my ideation prompt, optimizing my summarization prompt, etc.
I switched from markitdown to anydoc as my default PDF to Markdown converter. AnyDoc handles tables and other kinds of structures in PDF much better #ForNow.

Questions I was asked

Week ending 30 Aug 2026

Question: How hard was it to adapt your personal AI workflow to work?
Answer: I didn’t try solving work problems. Rather, I took what I was solving personally with agents and tried finding where at work I can apply it. “I have a hammer, let me find all the nails.”
Question: If agents can do the coding, how do we teach the foundational blocks?
Answer: We don’t care if they learn FastAPI, etc. Routing, authentication, interfaces, … - THAT is more useful. So give them tasks that forces them to learn FUTURE foundational blocks FROM agents.
Question: How do you monitor whether people are actually using reusable AI skills?
Answer: Read the agent session logs. A small script over Claude, Codex and Copilot histories can tell you which skills were actually used and how often.
Question: How are you determining which AI spend is dumb?
Answer: Start with the highest-cost users and inspect the obvious outliers. Give the raw logs to an agent, ask why someone spent that much, tell users to make obvious optimizations, etc. before attempting sophisticated optimization.

Mistakes I made

Week ending 30 Aug 2026

I said Roger Federer “picked up tennis at the age of 20 or something.”
Correction: Federer began playing tennis at age 8. The point in Range is that he sampled several sports and specialized later than Tiger Woods, not that he began tennis as an adult. Evidence: ATP Tour — Roger Federer biography
LOW · FALSE
I said a $20 ChatGPT subscriber can use Pi with GPT-4o without API-token billing, and implied OpenAI subscription usage effectively doesn’t rate-limit.
Correction: Pi does support ChatGPT subscription OAuth, but specifically through its ChatGPT Plus/Pro (Codex) provider. Pi lists GPT-4o under the regular OpenAI API-key provider, and Codex subscription usage has plan limits. Evidence: Pi — Providers Pi — GPT-4o model configuration OpenAI — Using Codex with your ChatGPT plan
MEDIUM · FALSE
I said the IITM Web-Enabled M.Tech AI credential “is not yet accepted by PhD programs, even in India, let alone abroad.”
Correction: I don’t have evidence for that blanket statement. IIT Madras formally awards the Web-Enabled M.Tech in AI, and IITM CSE’s PhD eligibility accepts M.E./M.Tech degrees in AI, ML and related engineering areas. Other universities make their own admissions decisions. I should say the program is relatively new and its research/PhD outcomes are not yet well established. Evidence: IIT Madras WSAI — Web-Enabled M.Tech in AI IIT Madras CSE — PhD eligibility
HIGH · UNSUPPORTED
I said Indian tax residency for NRIs comes down to whether you were in India more than 120 days.
Correction: The 120-day threshold is a special case, not the general rule. The normal tests include 182 days and 60+365 days; for an Indian citizen/PIO visiting India with more than ₹15 lakh of non-foreign income, 120+365 can apply. There is also a deemed-residency rule. Evidence: Income Tax Department — Non-Resident FAQs
HIGH · OVERSTATED

My Top 5 Prompts in August 2026

Fri, 28 Aug 2026 14:52:34 +0800

I save prompts and prompt fragments I regularly use with ChatGPT, Claude, etc.

(Prompt fragments are just prompts used along with other prompts. They’re typically smaller. But the difference isn’t important or anything… I just use two methods.)

I use a script triggered by Ctrl Alt P to select the prompt to paste.

This month, the five prompts / fragments I used the most were:

#5: Reframe question skill.

Sometimes, I’m not sure I’m asking the right question. Actually, I’m not even sure what I’m asking.

“Reframe question” roughly says, “Guess what I really need, say it, then answer.” The “say it” part is very helpful - I find out what I really meant to ask (or correct it.)

This is actually a skill, but since ChatGPT Plus does not yet automatically load skills in the “Chat” mode, I need to paste this manually.

#4: Compare models

I ask the same question to ChatGPT and Claude (and sometimes Gemini) and ask for a second opinion. That way, I get the best of both models, more thinking, and a sense of which models are good for what. As of now, I prefer:

ChatGPT: For analytics, rigor, algorithms
Claude: For strategy, soulful writing, creativity, front-end code
Gemini: For learning, readable writing, foreign language, people search

#3: Comic strip + Comic page

I paste my blog posts, talk transcripts, etc. and ask it for a single panel or full page summary. The panels are usually funny. The pages aren’t too informative but they’re usually engaging.

#2: Email Reply

Most of my email replies are based on this prompt. For over 80% of my emails, I just send its response as-is, and for about 15%, I send it with minor tweaks.

(To be fair, I wouldn’t have bothered replying to many emails earier, so the percentage I need to correct seems smaller than it really is.)

#1: Meeting transcript context

This is my top prompt. It creates a prompt to transcribe meeting recordings.

Now, that’s a weird thing to do, but here’s the situation.

Google Meet (and Teams) have poor transcripts. Whisper, Gemini, and most other models are much better. Here’s how I use Gemini.
But Gemini (like other multimodal models) doesn’t label speakers well - it doesn’t know who said what and gets names wrong sometimes.
Giving Gemini a hint about who said what works quite well. After some testing, I hit upon this prompt, which says, “Give examples of who said what so that a good model can label speakers.”

I run this after almost every meeting, so understandably, I use this a lot.

Here’s the usage count in August:

Count	Prompt
55	Meeting transcript context
47	Email Reply
19	Comic strip + Comic page
16	Compare models
11	Reframe question

Things I Learned - 23 Aug 2026

Sun, 23 Aug 2026 00:00:00 +0000

This week, I learned:

DuckDB 2.0 adds a CONNECT command that can connect to databases like MySQL, PostgreSQL, etc. making DuckDB the only DB client I need.
EQ-Bench evaluates models on capabilities like: does it follow direction, does it challenge you, how good are its insights, does it build rapport, etc. Very interesting to see that the Gemini models are the most “yielding” to your pressure and “validating” your beliefs (Anthropic’s are the least) while OpenAI models are the most “directive” (give concrete actions) #ForNow. There are other benchmarks such as Creative writing which Opus 5, Kimi K3, and GPT-5.6 Sol lead #ForNow.
OpenRouter offers several models at a discount. #ForNow, GPT-5.6 Sol is at a 50% discount, DeepSeek v4 Pro at 62%, and Gemini 3.7 Flash at 75% discount. There’s also a Free Models collection that #ForNow includes Nemotron 3 Ultra and more.
For a few years, I’ve been feeling useless, that I don’t contribute anything tangible to my organization. No measurable metric I’ve improved. Today, it strikes me that this is a good thing if I don’t want to be fired. As AI eats up more of our work, measurable contributions naturally shrink (AI does more, you do less/different work), and the vague “Oh, he’s probably doing some good” is a safer bet than “He contributed 10% to this metric last year, this year it’s 1%, can we justify his cost?” (I’m sure marketers will come up with a good term to cover this feeling of uselessness that is actually a good thing.)
ChatGPT Desktop - Work is a layer on top of Codex #ForNow (which I sort-of expected, but the session logs confirm this). It adds instructions that cover:
- Memory: from memory_summary.md, MEMORY.md, rollout summaries, and saved skill notes. Recheck decaying ones, mention if unverified.
- Folders: Temo work in work/, final in outputs/, local files use absolute paths.
- Coordination: How to start, fork, inspect, message, wait for, rename, … Codex tasks, how to use subagents.
- Automations: Available tools for reminders, schedules, monitors, follow-ups, and wake-ups.
- Knowledge management: known project → memory; specialist task → skill; external object → connector; subtask → subagent; recurring work → automation; finished artifact → Work UI primitive.
- Presentation: Use shell/scripts internally but hide it, describe outcomes in user terms.
- Apps/Connectors: Gmail, Drive, GitHub, Dropbox, etc.
- Skills: via SKILL.md
Neither ChatGPT Work nor Claude Work can read the ChatGPT / Claude chat conversations. But the chat conversations can access past conversations via “Memory”. That’s a pity, and one of the reasons I’m more often on “chat” than on “work” - it can refer to my past chats automatically, which helps build a kind of unstructured knowledge base. The other reason is that, at least on ChatGPT, chat does not consume usage limits #ForNow. ChatGPT work and Claude - both chat and work - consume usage limits.
Weird that there’s a “make a lot of money” button and nobody’s pressing it (take your SaaS, make it headless, let agents use it, charge per interaction esp for enterprises). Thariq
AI is accelerating discoveries in cyber (definitely) and maths (reasonably) but not as much in algorithms. METR
“Match your prompt style to the desired output.” Clear guidance from Anthropic that “The formatting style used in your prompt may influence Claude’s response style.” OpenAI says something similar - adapting implicitly to the user’s tone. But this is not a very strong signal - examples are better guides.
Why model routing must be in the harness. Makes perfect sense. “Only the harness can judge when a model switch is worth the cache miss.” I’m sure some popular harness (like OpenCode, Codex, Claude Code) will enable an “auto model” mode that’ll pick and change the model by itself by the end of the year.
Microsoft Print to PDF can, sometimes, generate PDFs with no highlightable or selectable text - all fonts get converted to paths. A crude solution is below. This is a poor solution but often good enough for an LLM to process. (Of course, if you’re passing it to an agent, you could just upload the file and it’ll figure it out.)
```
sudo apt install ocrmypdf tesseract-ocr
ocrmypdf --output-type pdf input.pdf ocr.pdf
pdftotext ocr.pdf -
```
When my train neighbor started talking to me (asking personal questions but was self-aware, rambling but was partly interesting), I asked if he was an extrovert. He said “No”. People who talk a lot can still be introverts if they’re:
- socially competent (like me at work)
- in “performance mode” (like me when I’m on stage)
- are high energy and engaged by topics (maybe him - or me when, like now, when I just HAVE to tell the flight attendant Ollama + Gemma 4 + Pi answering a psychology question is a delight!)
- ambiverts (maybe him)
- not self-aware and are mistaken (maybe him)
ffmpeg can embed subtitles. ffmpeg -i video.webm -i subtitles.srt -map 0:v -map 0:a? -map 1:0 -c:v copy -c:a copy -c:s srt -metadata:s:s:0 language=eng -metadata:s:s:0 title="English" -disposition:s:0 default output.mkv adds subtitles.srt to video.webm and creates output.mkv with embedded subtitles. Note: On VLC, MKV works better than WEBM if you want to embed subtitles. On the browser, you need to use the <video> tag with a <track> tag to display subtitles.
ffmpeg can burn subtitles. ffmpeg -i video.webm -vf "subtitles=subtitles.srt" -c:v libvpx-vp9 -crf 30 -b:v 0 -c:a copy output.webm re-encodes the video with subtitles added to the video.
ffmpeg can offset subtitles. For example: ffmpeg -itsoffset 10 -i input.srt -c copy output.srt creates output.srt with subtitles starting 10 seconds later than input.srt.
When I hear my father’s tales from his childhood, I’m struck by how much India moved forward in half a century on child mortality, consumerism, and communication (mobiles). Also surprising are what feels the same: legal system, travel (trains made it easy), food (tasted better, then), entertainment (theatres made it easy), gardening, education (scholarships made international study more accessible than I thought),
I asked ChatGPT how I adapt my message based on the audience. It discovered that I tailor messages to the audience’s (A) Objectives (B) Examples (C) Expertise - e.g. tell vs ask (D) Risk appetite. But what’s distinctive is that I often surrender, i.e. I don’t defend my view, but drop it and run with their framing.
“People with taste are picky. The only way to make money is by satisfying those who can’t discern quality.” Adrian Hanft. I’ve been telling people that taste is our differentiator against AI. And yes, it’s a fickle differentiator. I mean, how do you build a taste that thousands or millions will adopt? Or… is taste marketed more often than organically adopted, in which case, persuasiveness matters more than taste? But either way, unless most people disagree with you, you’re building conformity, not taste.
On a flight, I tried ollama launch pi --model gemma4:e4b-it-qat. It’s a reasonably sensible model. Power consumption is high, though. I was at about 8 watts with ~7 hours of battery life. While running, power spiked to ~50W (1.5h) and settled down to ~12W (4h) when idle. The llama-server process consumes some CPU/GPU even when idle, but I couldn’t get it back to the ~8W even after ollama stop. (It eventually did return to 8W after an hour. Not sure why.) When I tried again on 21 Aug, it went up from 8W (8h life) to 28W (3h life) and back to 8W, so looks like when idle, it doesn’t consume power. I look forward to using local LLMs more!
Pain is good. Struggle is good. Stretch is good. Not new. But worth reminding, worth seeking.
T3 Code is a coding agent orchestrator. It lets you “remote control” multiple coding agent sessions across systems. The ecosystem of tools around coding agents is growing. Observability, e.g. AgentsView, is one such area.

Questions I was asked

Week ending 23 Aug 2026

Question: How do I safely let an agent modify my files when it could get things wrong?
Answer: Make backups and let it work on a copy. Try it five or ten times; once it repeatedly earns your trust, gradually remove the safety net.
Question: When should I turn an ad-hoc agent workflow into automation?
Answer: If I run it once every two months, I don’t mind the agent writing the code again. If it’s every two weeks or two days, save the script and automate it.
Question: How do we decide which agents to train when client problems keep changing?
Answer: Decouple the agent from the skill. Keep the skill intelligence agnostic - it’s not about correcting agent errors but about transferring context it won’t have. Keep a central folder of skills with one-line descriptions; whichever agent people use can scan it and pick the relevant skills.
Question: Which part of the current agentic AI narrative is overhyped?
Answer: GraphRAG is definitely overhyped. Prompt engineering is outdated; harnesses and agentic loops are not overhyped yet.
Question: To make agentic software development scalable, do we need a standard framework or just give everyone Cursor and let them figure it out?
Answer: Install Cursor for everyone and let them figure it out. Share lightweight enterprise guidelines as skills, but give every instruction an expiry date and a small benchmark so you can remove it as agents learn to handle it themselves.
Question: If frontier video models fail on physics and action scenes, how should we fine-tune them with our proprietary video data?
Answer: Don’t solve the physics problem; solve a much narrower action-block problem. Build reusable filtering and fine-tuning pipelines so the next frontier model can replace the base model and you train only on what it still cannot do; no manual annotations.
Question: Will AI take all the tech jobs in the next five to six years?
Answer: Yes. And so what? AI will take a significant number of existing jobs, and we’ll create new ones because our desires and competition don’t disappear; figure out which new work takes you further before your neighbor does.
Question: How did you come up with this conceptual clarity about what to do?
Answer: I didn’t. Pretend to have clarity, ask AI everything and use its answers, then do it so often and fail repeatedly that you get a feel for what works. Quantity beats quality like crazy.

Mistakes I made

Week ending 23 Aug 2026

I said Claude Code auto mode made the risk of unintended actions “negligible,” and later said “as of this month, it won’t make a mistake” with a “90% chance” it would preserve undoability.
Correction: I was far too confident and invented a probability I could not support. Agent safeguards are probabilistic. Anthropic’s own auto-mode evaluation reported a 17% false-negative rate on real “overeager” dangerous actions. For destructive local operations I should still use backups/version control, limit permissions and retain review where the blast radius matters. Evidence: Anthropic — How we built Claude Code auto mode
HIGH · OVERSTATED
I said “there is no difference between Codex and ChatGPT Work” and that Work is essentially Code with a lighter, more marketable name.
Correction: Work uses Codex technology, so the overlap is real, but they are distinct experiences. Work is aimed at longer multi-step research, analysis and deliverables; Codex remains specialized for software development and has separate workflows/history. Evidence: OpenAI — Introducing ChatGPT Work OpenAI — ChatGPT Work and Codex
MEDIUM · OVERSTATED
I said “1960s is when Studio Ghibli starts trying to catch up” with Disney.
Correction: Studio Ghibli was established in 1985. Evidence: Studio Ghibli — company history
LOW · FALSE
I said “I’m yet to find a use case where fine-tuning is worth it … Zero. Out of thousands of use cases … it was never worth it” and “in neither case is fine-tuning useful.”
Correction: “I haven’t personally found an ROI-positive fine-tuning case yet” would have been defensible; “never useful” is not. Fine-tuning remains a standard adaptation technique even for open-weight models; Meta’s official Llama cookbook includes fine-tuning and parameter-efficient fine-tuning recipes. Evidence: Meta — official Llama Cookbook Meta — Llama fine-tuning overview
HIGH · OVERSTATED
I referred to “ChatGPT’s share price” falling if a privacy controversy were real.
Correction: OpenAI was not publicly traded, so there was no public ChatGPT/OpenAI share price to fall. I should have referred to OpenAI’s private-market valuation, tender/share price, investor appetite or commercial impact. Evidence: Reuters — OpenAI’s planned IPO
MEDIUM · FALSE
I said Meta’s glasses can record someone “without you getting even an inkling or a notification.”
Correction: Meta’s AI glasses have an outward-facing capture LED that blinks while photos or video are being captured; current models disable the camera if the LED is covered or disabled. The indicator may be easy to miss, but there is one. Evidence: Meta — AI glasses privacy and capture LED
MEDIUM · FALSE
I said Alexa sends audio to Amazon only if you say “Alexa.”
Correction: The wake word is the normal trigger, but it is not an absolute rule. Follow-Up Mode allows requests without repeating the wake word, and Amazon says Alexa can sometimes mistake unrelated speech for a follow-up request. Evidence: Amazon — Alexa Follow-Up Mode Amazon — Alexa FAQs
MEDIUM · OVERSTATED
I said “In India, most children are born on 1st June, which is the admission cutoff date for most schools.”
Correction: There is a real historical June-1 anomaly in recorded dates of birth in parts of India: when exact birth dates were unknown, some schools/parents used June 1 for admission records. That does not mean most Indian children are actually born on June 1, and the cutoff is not universal nationwide. Evidence: Times of India — “Admit it, June 1 isn’t your real b’day”
LOW · OVERSTATED

What to do less because of AI

Fri, 21 Aug 2026 15:33:42 +0800

Some ideas on what we can do less, thanks to AI.

Why? Because less is more and we keep getting advice on what to do more of, not what to do less of.

Skip meetings you don’t speak in. Agents can summarize the transcript.
Stop taking notes. Agents can summarize a recording.
Stop reading group chats. Agents can summarize WhatsApp groups, Hacker News, Twitter, and more.
Stop translations. Agents translate on the fly. For you, and your audience.
Stop weekly status reports. Agents can tell you (and your manager and team) the status.
Stop building dashboards. Agents tell you “what’s wrong” or “what’s next” directly.
Stop building software at first. Tell the agent to do the work. Let it write software if needed. Save it if you use it often.
Stop writing requirements. Agents can prototype. Test if that’s what you need, first.
Stop teaching coding. Agents can code. Teach people to use agents.

Also - a few things that were true pre-AI:

Stop traveling unless handshakes matter. Actually, this was true even with Zoom, etc.
Stop live lectures unless there’s interaction. Actually, this was true even with YouTube, etc.

Bonus:

Stop trying to remove Jordan’s third hand. You can’t. You just can’t. He is jinxed. (For now.)

Local agents are good but slow

Fri, 21 Aug 2026 12:40:28 +0800

Last year, I shared how local LLMs are expensive. There are only 3 reasons to use them over an inference API:

You have no choice, e.g. Government, pharma, finance, or other sensitive data.
You have free GPUs, e.g. your laptop, local data center with sunk cost, free cloud credits, etc. This doesn’t scale.
You want to learn.

This year, I finally started using them for a combination of #1 (on flights) and #2 (my 8 GB NVIDIA RTX 2000).

It began with simple queries on Edge Gallery on my mobile - like “Is this vegetarian?”, “What do you call those caption boxes in comics?”, etc.

I recently read how easily you can run Pi with Ollama and that ornith:9b and gemma4:e4b-it-qat are good models for my system so I can just run:

ollama launch pi --model gemma4:e4b-it-qat

… to get a reasonable coding model locally.

I tried it for a few non-technical things first. For example:

That morning, I asked my talkative train neighbor, “Are you an extrovert?” He said “No, I’m an introvert.” On the flight, Gemma4 explained the contradiction well
I tried reducing my battery usage on the flight by asking it what different processes were. It didn’t solve the problem but gave me some ideas.

But that’s not to say it’s great. For example, when I asked it:

Can you find a blog post where I spoke about an incident at BCG in the early 2000s (I may have written about it later) where I asked them, “you keep saying we should do more of this and more of that, what should we do less of”? I think there is an entire blog post about that.

Here’s the approach it took (comments are mine):

# List my blog posts - which showed this: https://github.com/sanand0/blog/tree/main/posts
ls ~/code/blog/posts/

# Search for "BCG" - which returned 63 results, NONE of which were the post I needed!
grep -r -i "BCG" ~/code/blog/posts/

# Read the most likely first candidate
cat ~/code/blog/posts/2000/another-day-at-lbs.md

# ...and the next
cat ~/code/blog/posts/2000/boston-consulting-group-interview.md

# .. and the next
path: ~/code/blog/posts/2012/the-next-chapter-of-my-life.md

# Back-track and search for doing more or less in 2000-2003 - again, no results
grep -r -iE "d o more|d o less|sh ould" ~/code/blog/posts/200[0-3]/

# Try across all years - this gives 2718 results
grep -r -iE "more|less" ~/code/blog/posts/

… and it’s been 20 minutes so far - but it probably would have gotten to the answer if it hadn’t crashed for some reason.

On the other hand, GPT-5.6 Sol (on ChatGPT) did the following (comments are mine):

# Write a couple of multi-line scripts to search for "do more", "do less", etc. AS WELL AS for "BCG" and "Boston Consulting" and "consult"
set -euo pipefail
BLOG="$HOME/code/blog/posts"
test -d "$BLOG"
echo '--- phrase matches ---'
rg -n -i --glob '*.md' --glob '*.html' --glob '*.txt' \
  -e 'do more' -e 'do less' -e 'less of' -e 'more of this' -e 'what should we do less' \
  "$BLOG" | head -120 || true

echo '--- BCG/context matches ---'
rg -n -i --glob '*.md' --glob '*.html' --glob '*.txt' \
  -e '\bBCG\b' -e 'Boston Consulting' -e 'consult' \
  "$BLOG" | head -120 || true

# That failed. It provided the wrong home directory: /home/oai/share
#   - It got confused with its own container.
# But it corrected itself and tried again. That got ~220 GOOD matches.

# Now, it does 3 things:
# 1. Read the most promising post (which is the correct one): https://www.s-anand.net/blog/less-is-more/
# 2. ALSO search in the file for related terms like Infosys, manager, office, consultant, years ago, etc.
# 3. A BRILLIANT step - search for other phrases, like "do less... do more" or "what... do less", etc.
#    Just in case it missed something!
set -euo pipefail
F="$HOME/code/blog/posts/2008/less-is-more.md"
echo '--- less-is-more.md ---'
nl -ba "$F" | sed -n '1,125p'

echo '--- nearby explicit provenance clues ---'
rg -n -i -C 3 'BCG|Infosys|manager|office|consultant|years ago|back then|firm' "$F" || true

echo '--- exact/near quote elsewhere ---'
rg -n -i --glob '*.md' -C 2 \
  -e 'you can.t do more of something unless you do less' \
  -e 'what.*do less' \
  -e 'do less.*do more' \
  "$HOME/code/blog/posts" | head -120 || true

That’s it. That’s all it needed to find the post.

What impressed me about both agents is that:

They used smart synonyms just from my query - they know how people typically mis-spell.
They corrected themselves as they went along. For example:
- Pi found no results between 2000-2003, and extended to all my blog posts.
- ChatGPT got the home directory wrong, but corrected itself.
They revised strategies as they learnt. For example:
- Pi realized I might not have mentioned “BCG” in my post.
- ChatGPT realized I might have multiple blog posts about this.
- Both updated search terms to use based on the results. For example, ChatGPT learnt I was at Infosys from the search results and added that as a search term!

So, I’ve reached a mental tipping point. Local agents are good enough in an emergency. They’re slow, but they can get my work done if needed.

What can AI not do

Thu, 20 Aug 2026 16:15:00 +0800

Over the last few weeks, I asked 166 people:

What’s something you did not ask AI last week, because it can’t do it?

There were 4 groups of responses.

37% said “Reliability”. AI tries but fails often enough. Like:

Verify & return exact source of information.
Generate accurate images.
Completing PDF accessibility reliably according to the WCAG 2.2 standard.
Converting 2D interior design to 3D view as vision models are not good keeping things as per scale.
Create a user manual based on my codebase. I have asked this before and it has failed.

27% spoke about “Reach”. AI can’t access tools, data, or do something.

Book my tickets for a vacation.
Update my to do list.
Open my application at 5pm, after execution, Open GChat and Send Files…
Read … API documentation that is not openly published on the internet.
Run a Business Objects Report…

Interestingly, agents are steadily improving on both the reach (with more tools and permissions) and reliability (with deterministic verification and better models).

14% said “It can do everything”

I think it has all the answers
I have used AI for all possible tasks
Nothing in fact, but decide whether to take its answer or not later. Anyways try it to see what it says.

12% argued “Presence”. Things that need interpersonal or physical presence.

Attend my meetings.
Have a difficult conversation with a colleague.
Talk to stakeholders for requirement gathering.
Doctor’s appointment.
Play … with me.
Eating.

That’s in order of decreasing frequency - but also increasing correctness, I think.

Reliability: model intelligence and agent reliability are consistently improving. I rarely worry about hallucinations.
Reach: connectors are growing and agents can use cards to buy stuff. I expect reach is fading faster than we think.
It can do everything: Well, it can probably replace the brain, like machines can replace the body, but maybe there’s more than that?
Presence: This feels like one of those things which, almost by definition, seems hard to delegate.

My prediction: By the end of the year, people will realize that agents are far more reliable, have far more reach, and the percentage citing these problems will fall below 25% and 15%. Presence will be the bigger problem AI cannot solve.

Pleasures of a bus ride

Wed, 19 Aug 2026 17:54:02 +0800

There are some small pleasures in life, not worth mentioning by themselves, but sometimes they come together so well. Like:

Google Pay popping up a “Disable verification for transit?” option, so I just touch my phone to the bus ticket reader even while on a call and it just works like my card. (It works even without Internet!)
Finding a seat at the front of the double-decker bus on top, so I can take in the entire vista from a height for the entire one hour ride
Discovering the ultra-configurable Open Camera app (thanks to Thej) and capturing at the much smaller 640x480 resolution at 10x speed
Transferring it like a breeze to my laptop with LocalSend
Compressing it into an ultra-small 1.4MB video thanks to the AV1 codec and FFmpeg
Publishing it to Cloudflare within a second thanks to ultra-fast WiFi

Watch the video

Yeah… this is life!

Watching videos with a phone case

Mon, 17 Aug 2026 18:19:49 +0530

Two months ago, I was impressed at the ingenuity of a lady who used a plastic cover to watch videos on a flight.

Today, on the same Indigo flight from Singapore to Chennai, I saw a man one seat ahead of me use the headrest cover on the seat in front to watch videos.

(The headrest cover is the cloth behind your head on airline seats. It saves cleaning effort.)

He took out his phone case, placed it behind the headrest cover, and pushed the phone back into the case. Now, the cover is caught between the phone and case and the phone just hangs from cover - which is quite firmly attached to the seat!

I asked him when & how he started doing this.

I have to hold it in my hand otherwise. It was my first time today.

This flight is filled with ingenious passengers. I learn a lot from them!

My predictions in 2025

Mon, 17 Aug 2026 09:20:33 +0530

In 2025, I made a number of predictions on this blog. (Not intentionally. I just said stuff.)

I asked ChatGPT to audit them. It selected 440 claims, filtered out vague or pending ones, and verified the rest. Here’s what I got right and wrong.

🟢 “My chat will overtake search in 12-18 months. When ChatGPT becomes my primary lens on knowledge…” (chatgpt-vs-google-usage.md:L27). Audit: This actually happened in your own browsing data: search led through April 2026; in May, chat jumped to 2,211 visits vs 1,319 search visits, and stayed comfortably ahead thereafter.
🟢 “Typed languages are better suited for vibe coding. This will likely lead to the growth of typed languages (TypeScript, Rust, Go) but also of typing in untyped languages (e.g. Python).” (things-i-learned-10-aug-2025.md:L33). Audit: TypeScript became GitHub’s #1 language in August 2025 and grew 66% YoY; GitHub itself explicitly connects the rise of typed languages to more reliable AI-assisted coding. This is unusually strong because you got both the direction and mechanism. (The GitHub Blog)
🟢 “Code Mode … is a smart way to use MCPs and a very likely future direction. Using LLMs to write code to call MCPs rather than directly.” (things-i-learned-05-oct-2025.md:L49). Audit: OpenAI’s current Responses API has essentially this as a named capability: Programmatic Tool Calling lets a model write and execute programs that coordinate multiple tools and intermediate results. (OpenAI)
🟢 “CLI optimization for LLMs will likely emerge. More CLIs (and wrappers / hooks in the shell) will improve output and error contexts for LLMs…” (things-i-learned-27-jul-2025.md:L86). Audit: By May 2026 you were yourself testing rtk, a CLI proxy explicitly producing compact agent-friendly command output; across 216 commands you measured about 50% token reduction. That is almost exactly the wrapper you predicted.
🟢 “In the future, AI that works directly with file systems, Model Context Protocols, and local APIs are likely to become more important.” (features-actually-used-in-an-llm-playground.md:L76). Audit: File-system-native coding agents became mainstream, while MCP expanded into both hosted and local integrations; Anthropic now packages local MCP servers as easy-to-install desktop extensions. (Claude Help Center)
🟢 “Agents are slow. Parallelizable tools … will grow. Tool speed … will become more important.” (things-i-learned-22-jun-2025.md:L106). Audit: Parallelism has become a central agent UX: OpenAI’s Codex app is explicitly built around managing multiple agents simultaneously; at the extreme, internal users now accumulate more than 60 hours of agent turns per day by parallel execution. (OpenAI)
🟢 “Companies-of-one will grow. Sole founder can handle support functions.” (things-i-learned-24-aug-2025.md:L14). Audit: Nasdaq’s Economic Institute finds one-person US business applications up more than 20% since early 2025, with essentially all recent application growth coming from solo businesses. Stripe separately reports solo founders reaching 63% of Atlas C-corps in Q2 2026. (Nasdaq)
🟢 “We will move towards an organization structure where developers are embedded with business teams rather than working as a separate group. Sort of like embedded executive assistance instead of a central typing pool.” (things-i-learned-08-jun-2025.md:L49). Audit: Forward-deployed engineer demand reportedly increased 42-fold from 2023-25, with roughly 9,000 roles globally by early 2026. The role is almost precisely “technical people embedded with the business to make AI work in its real environment.” (Reuters)
🟢 “Shadow apps will grow. Anyone can code. Users build apps with prompts, sheets, agents, outside of IT SDLC. Like Excel sheets.” (things-i-learned-24-aug-2025.md:L18). Audit: Microsoft now explicitly describes a “new wave of shadow AI”: users installing coding/desktop/SaaS agents outside traditional IT governance, and has built discovery products specifically for unmanaged AI applications and agents. (Microsoft)
🟡 “Agents generate diffs/PRs. Tools to edit and comment on these online will emerge.” (things-i-learned-22-jun-2025.md:L107). Audit: GitHub now measures PRs created and merged by Copilot coding agent, while review comments can be handed directly to the agent with “Fix with Copilot,” including batches of review feedback. That’s almost verbatim fulfillment. (The GitHub Blog)
🟡 “Models’ ability to orchestrate longer workflows will improve. Factor that into your application design.” (things-i-learned-10-aug-2025.md:L44). Audit: By mid-2026, OpenAI reports large increases in requests corresponding to >30-minute, >1-hour and even >8-hour human tasks, while Codex explicitly targets long-running tasks spanning hours or longer. (OpenAI)
🟡 “Code review process will be re-invented.” (things-i-learned-22-jun-2025.md:L109). Audit: GitHub has rebuilt Copilot review around an agentic architecture that gathers broader repository context, uses tools, produces findings, and can hand fixes to another coding agent. This is substantially more than autocomplete added to old review. (The GitHub Blog)
🟡 “Domain expertise will therefore become even more valuable in the near future.” (things-i-learned-20-apr-2025.md:L39). Audit: 2026 hiring evidence points toward domain/product expertise becoming more important rather than pure coding alone, particularly as AI handles more implementation and firms need people who can connect it to actual business functions. (Reuters)
🟡 “Validation is the New Bottleneck: Since coding is now much faster, the critical, time-consuming task has shifted to reviewing, testing, and validating the LLM’s output.” (things-i-learned-17-aug-2025.md:L90); you also predicted “The Quality Control (QC) function will become larger and more critical” (L95). Audit: GitHub has now productized exactly that bottleneck in Code Quality; more than 10,000 enterprises used its preview, and GitHub explicitly frames AI-accelerated code output as creating the need for trustworthy pre-merge quality validation. (The GitHub Blog)
🟡 “Agents generate technical debt faster than humans. Solving this will become a major problem/opportunity.” (things-i-learned-22-jun-2025.md:L114). Audit: GitHub’s 2026 Code Quality launch is close to a commercial instantiation of this forecast: AI increases code output, so automated quality/debt detection and remediation moves earlier into the development cycle. (The GitHub Blog)
🟡 “Governance will grow. Non-experts are acting like experts. Validation is more important.” (things-i-learned-24-aug-2025.md:L19). Audit: The companion to shadow AI has indeed been governance: Microsoft now ships specific discovery, monitoring and governance for unmanaged AI agents, while NIST has continued expanding formal GenAI evaluation tooling. (Microsoft Learn)
🟡 “Soon, we won’t just follow a lesson plan – we’ll have lessons built just for us. AI will track how we learn and adapt in real time. It’ll feel like having a personal coach in your back pocket.” (o3-is-now-my-personalized-learning-coach.md:L91). Audit: ChatGPT Study Mode now asks what the learner knows, adapts explanations, checks understanding, works from uploaded course material, and uses memory to personalize support; OpenAI explicitly describes the objective as personalized learning support available to any student. (OpenAI Help Center)
🟡 “Cost is going down so quickly right now that all you have to do is wait, and stuff will become available for a very affordable or even a free price.” (things-i-learned-16-mar-2025.md:L116). Audit: The broad direction held. OpenAI cut GPT-5.6 Luna API prices by 80% in July 2026 while simultaneously improving capability-per-dollar. The “all you have to do” part is hyperbole, but the price-curve forecast was right. (OpenAI)
🟡 relayed: “Control of chips and GPU compute is what will likely be the gameplay to control AI dominance globally.” (things-i-learned-02-feb-2025.md:L16, attributed there to Dario Amodei). Audit: Advanced-AI-chip export licensing remains an explicit geopolitical control mechanism in 2026, including restrictions and license review for H200/MI325X-class accelerators going to China.
🔴 “AI closes the gap between junior & senior devs – even when both use AI. Quality doesn’t suffer much. So onboarding can be faster, compensation ladder may shorten.” (things-i-learned-03-aug-2025.md:L52). Audit: The emerging evidence says AI changes the work but does not erase the expertise gap: experienced developers are better at steering/delegation, while low-experience AI-heavy contributions incur substantially more review and lower acceptance. (arXiv)
🔴 “LLMs already deliver hours of analyst work in minutes. Entry-level roles WILL vanish.” (goodbye-mba-hello-ai.md:L17). Audit: The labor-market warning was directionally good, but “vanish” is a major magnitude error. Stanford finds a meaningful relative decline among 22-25-year-olds in highly AI-exposed jobs, while employment remains substantial and overall exposure groups still show employment growth. “Entry-level hiring contracts sharply” would have scored much better. (Stanford Digital Economy Lab)
🔴 “Coders micro-manage LLMs. I think a novice will be more efficient and get better results than me.” (how-to-visualize-data-stories-with-ai-lessons.md:L285). Audit: Current empirical work points the other way in real software work. In a 22,953-PR study, lower-experience AI-heavy developers received 4.5* more review comments, had 31% lower acceptance, and took over 5* longer to resolve issues; qualitative work likewise finds experienced developers better at delegation and control. (arXiv)
🔴 relayed: “API access from model providers will shrink. Selling tokens is not a viable business model given lowering costs.” (things-i-learned-23-mar-2025.md:L19, from the Alexander Doria notes immediately above it). Audit: Almost exactly backwards. Model providers expanded their APIs into richer agent platforms, and token-metered API access remains a core commercial model - including premium pay-as-you-go modes. (OpenAI)
🔴 “APIs are likely to be replaced by just chat requests that will do the same thing. APIs might be replaced by RPA, where somebody uses a chatbot to do the equivalence instead.” (things-i-learned-16-mar-2025.md:L111-L112). Audit: Chat did become a front end, but the implementation moved toward more APIs underneath, not fewer: tool APIs, Responses, MCP, computer-use interfaces and programmatic tool calling are now the substrate agents use. (OpenAI)
🔴 “Software companies build ‘SaaS’-like apps today. Agents will replace apps. Instead of UI, workflows, and app logic, they’ll engineer prompts, APIs, and evals.” (agents-will-replace-saas-apps.md:L12). Audit: The interface-shift was right; “replace” was not. Gartner now forecasts agentic AI may expose roughly 20% of SaaS application spending by 2030 - meaning substantial disruption, not app extinction. Agents are often a new interaction layer over systems of record and APIs. (Gartner)
🔴 relayed: Models will “internalis[e] workflows … to wipe out the apps and workflow space.” (things-i-learned-23-mar-2025.md:L17, from Alexander Doria notes). Audit: “Internalize capabilities” was insightful; “wipe out” was the failed extrapolation. Enterprise applications remain a very large substrate even in Gartner’s fairly aggressive agentic-AI forecast. (Gartner)
🔴 “Demand for SaaS (one-size-fits-all) will shrink.” (things-i-learned-06-apr-2025.md:L89). Audit: Not yet. For example, Gartner forecasts Indian SaaS spending growing 18.9% in 2026, from $3.9B to $4.6B. AI is changing SaaS economics and seat licensing, but current demand is still growing rather than shrinking. (Gartner)
🔴 “The early majority have come in… Soon the late majority will come in asking for existing solutions that have already solved their problem for many others.” (things-i-learned-03-aug-2025.md:L44). Audit: This mapped your client/audience experience onto population adoption much too quickly. In 2026, US Census data put business AI usage around 17-20%, nowhere near a conventional late-majority phase. (Census.gov)
🔴- baseline error: “Given the cost and accessibility of drones, I guess drone terrorist attacks will soon emerge.” (things-i-learned-16-nov-2025.md:L43). Audit: They had already emerged. The UK government documented Daesh using small armed remotely piloted aircraft carrying grenades in Iraq in 2017; the UN had already been studying weaponized UAS use by non-state armed groups for terrorism-related purposes before this 2025 post. This is therefore a clean failure to establish the baseline, not a future hit. (GOV.UK)
🔴 relayed: “Personal writing with connection won’t go away. AI can’t give you heartbreak. But the rest of non fiction writing will vanish.” (things-i-learned-30-mar-2025.md:L69, under “Notes from Writing with AI”). Audit: Nonfiction is under real pressure, but “vanish” is nowhere close. UK nonfiction publishing still generated about GBP1.0B in 2025, down only 3%; the wider publishing industry reached record revenue. (Publishers Association)

Legend:

🟢 = I got it right
🟡 = I might be right?
🔴 = I got it wrong

I was wrong when making extreme predictions, e.g. APIs, SaaS, workflows, junior devs will be replaced by agents. They did shrink (so I was directionally right) but they’re far from replaced.

I was right when making directional predictions, e.g. CLIs, shadow apps, verification, solo founders, … will grow. These are easier and safer to predict (hence less impressive).

There’s a meta-insight here. You can take anyone’s corpus (meeting transcripts, emails, blog, books, research) and see how good their predictions are. This makes auditability for more accessible. So, I predict that we’ll see new roles & business opportunities emerge in 2026 based on this. For example:

Auditing researchers’ predictions
Auditing analysts and industry experts’ claims
Auditing company forecasts and promises
Auditing leaders’ promises
etc.

Things I Learned - 16 Aug 2026

Sun, 16 Aug 2026 00:00:00 +0000

This week, I learned:

psub is a neat fish option to treat command outputs like file inputs. E.g. diff (sort a.txt | psub) (sort b.txt | psub)
More anchor points on how much text to ask AI for:
- … in 200-300 words (about 1 book page)
- … in ~1 A4 sheet (~500 words)
- … in ~3 minute of conversational Grade 8 reading (~500 words)
I usually log things in a single file rather than split things into yearly, monthly, etc. A single file is more portable, scannable, and, for human logs, gives a feeling of accomplishment. I split when the file gets so unwieldy that it actually hurts - which is a good thing because it means I’m actually using it (maybe?) - and splitting beforehand may be good planning but is also premature optimization, adding friction to a nascent, fragile process. (This idea is pretty generalizable.)
Claude models launched after 2 Aug 2026 “weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.” How Claude marks AI-generated content.
VS Code has built-in dictation. It uses nemotron-3.5-asr-streaming-0.6b by default #ForNow, and it’s pretty good. I expect I’ll use it a lot more, since most of my typing is in VS Code anyway. Ctrl + Alt + V toggles dictation. Dictation is also supported in the terminal and is trained to recogni
I had an interesting moment today when I asked ChatGPT to identify which of my ChatGPT conversations were the most effective. I gave it access to my computer and it started using the browser to scrape itself so aggressively that I stopped it from it banned itself!
GeoLibre looks like a full-fledged GIS. It’s open source and runs directly in the browser.
When installing tools with mise, if it messes up the platform, you can explicitly specify it. For example: mise use -g 'github:pranshuparmar/witr[asset_pattern=witr-linux-amd64,bin=witr]@latest'. Mise Docs
I saw this snippet from Claude Code: await Promise.race([document.fonts.ready, new Promise((r) => setTimeout(r, 3000))]); That waits for up to 3 seconds for the fonts to load. Didn’t know about Promise.race() before. It’s different from Promise.any() in that .race() will return the first result while .any() will return the first success.
Measure the verification / follow-up tax in your AI work. AI does things fast. You have to take the next step. It helps to tag items with how long they’d take to verify or action. We do for specification, because we have to do that now, before telling AI. But I typically ignore the deferred tax.
A theme is emerging: “Human attention needs reciprocity. If you ask for someone’s attention, first demonstrate your own effort.” Example: “people really don’t like when a coworker’s chatgpt contacts them asking for help with a task, even when they’d be perfectly happy doing that same work if asked by that coworker.” Greg Brockman

Questions I was asked

Week ending 16 Aug 2026

Question: What should a strong data scientist actually build in an AI-native delivery model?
Answer: Assume he is training an AI to replace him. He should not build the thing himself; he should direct the agent, apply his judgment over a few iterations, and leave behind a portable system you can benchmark and rebuild simpler, better and faster.
Question: If two people are iterating on an AI-native delivery workflow, what collaboration setup do they need?
Answer: Shared files solve most of it. Keep data, code, skills/prompts and notes in folders with the right permissions; Google Drive or OneDrive is enough for now.
Question: If we’re debating whether a course is even necessary anymore, how do we transition it for AI?
Answer: Start with a single class that’s full AI, a single exam that’s full AI, one step at a time.
Question: Instead of hiring more developers, should I take fewer people and spend the difference on premium AI seats?
Answer: Experiment with a few people, not everyone. Treat AI as an extra headcount slot, but radically raise the output expected; getting that productivity happens only when you need that productivity.
Question: Should we self-host open models to reduce LLM costs?
Answer: Do the math: machine cost per hour versus useful inferences per hour and compare it with the API. Cost alone probably isn’t enough; privacy is a much better reason to self-host.
Question: How do you build a QA agent that tests developers’ work and reports back?
Answer: Don’t build the agent first. Tell a coding harness the outcome—find requirements, create tests, run them and report—and do it manually 5–10 times; automate only after you know what you actually want.

Mistakes I made

Week ending 16 Aug 2026

I said “The data is secure everywhere. All of them have solid enterprise contracts” and that choosing an AI tool for company data was not really a technical question.
Correction: Enterprise AI offerings can have strong security, but their data handling is not interchangeable. It depends on the exact product, plan, tenant configuration, retention and training settings, connectors, geography and contract. Microsoft explicitly says Copilot controls vary by subscription; OpenAI and Anthropic make separate commitments for their business/commercial products. I should check the approved product and its actual controls rather than assume equivalence. Evidence: Microsoft — Copilot enterprise data protection OpenAI — Enterprise privacy Anthropic — commercial-product training policy
HIGH · OVERSTATED
I said GitHub is where you take version-controlled software and “save it publicly.”
Correction: GitHub repositories can be public or private; GitHub Enterprise also supports internal repositories. Version control does not imply publishing the code. When explaining this to a beginner, I should explicitly distinguish Git from GitHub and repository visibility. Evidence: GitHub Docs — About repositories
LOW · FALSE
I said animation “started” in the 1930s with Snow White and the Seven Dwarfs.
Correction: Snow White was a landmark, but animated features predate it. El Apóstol was released in 1917, and The Adventures of Prince Achmed from 1926 is the earliest surviving animated feature. My visualization was a filtered view of popular IMDb titles, not a history of when animation began. Evidence: BFI — animated features before Snow White
LOW · FALSE

Should we fix poor design

Fri, 14 Aug 2026 10:09:23 +0530

Sugar (or condensed milk) in tea I get often stays unmixed at the bottom. I forget to stir it.

Is that poor design that should be fixed?
Or is the steadily increasing sweetness of the tea a brilliant feature?

Anyway, I therefore lick the bottom of the cup to get every last bit of sweetness.
But this cup is so deep that my tongue can just about reach the bottom.
And I can’t reach the bottom near the handle, because my nose is in the way.

Is that poor design that should be fixed?
Or is the 3 minutes of delight I get from fighting with the cup, making it a game, a brilliant feature?

Now that I’ve had it, I probably shouldn’t take sweets until the evening.

Is that agony poor design that should be fixed?
Or is the delight of one chunk of chocolate after a day of abstinence and yearning a brilliant feature?

(My brain feels there’s some cool insight here, but can’t figure it out. Is that poor design? Or our best differentiator from AI?)

Ask AI Anything Email

Mon, 10 Aug 2026 09:57:31 +0530

I ran, an experiment in Straive. I asked my colleagues to mail me any questions or tasks that they had. I’d use my agents + my knowledge as context and reply based on that.

A few interesting things came out of it.

First: It often does a better job than me - it has more patience and speed. For example:

One colleague sent a detailed pitch. My agent spotted a tiny arithmetic averaging error.
Another asked if I knew an ontology SME. The agent found a perfect match in my contacts - someone I would not have thought of (and reached out to.)
Another asked why a slide (screenshot) was empty. The agent searched Google Drive, found the slide, analyzed it, and replied: “It’s white-on-white text. But don’t bother. The slide two pages earlier is better for your meeting!”

One colleague asked: “Instead of sending this to your email id, why can’t we ask this directly into Claude or so?” It replied:

The difference is not the model. It is the context and tools. This agent can search my email, calendar, meeting transcripts, Drive, notes, demos and code, then answer using my recent decisions. A plain Claude, ChatGPT or LLMFoundry session can do the same only if it is given or connected to those sources.

… and then listed 3 answers that relevant to him that only I could give!

Where it does worse is, perhaps 2–5% of the time, I read the answer and think, “WTF? You’ve missed the whole point.” Maybe it missed the politics / psychology / nuance. But since I disagree with myself and can’t even remember what I wrote some of this might be noise.

Second: I went to Inbox Zero! Once I got there, it’s actually a slightly lonely feeling (“Oh, no one wants to talk to me any more!”) so I’ve started reaching out to people, asking them to mail me so that I can have AI reply to them.

So, here’s an offer:

Until 31 Aug 2026, you may email askai@s-anand.net. My AI agent (with my knowledge, code and tools) will reply within 24 hours.

This is best for questions where:

AI won’t know some things (e.g. based on my research / experiments / classes)
I wouldn’t have replied to you (e.g. I’m busy, or the question takes too much effort)

Examples:

What’s the most relevant thing you’ve already written, built, taught, or seen for this problem?
Given everything you know about this situation, what would you do?
You said X a year ago. Do you still believe it? What has changed?
What have you learned about X from your recent experiments, classes, or client conversations?
I’m pitching X to Y. Based on what you’ve seen work, what would you change?
Who do you know who might be unusually good for X?
Here’s my deck / proposal / code / model. What am I missing? What would you do differently?
This is going to take me three hours to figure out. Can your agent figure it out instead?

The best requests are probably ones where asking plain ChatGPT would give you a reasonable generic answer, but something I know, have done, or have access to could substantially change that answer.

Long, detailed emails are fine. Please attach/forward the context (document, email, links, screenshots, …). Short, 1-line questions are fine, too.

I batch runs daily, so please expect a response within about 24 hours.

Dancing with eyes closed

Sun, 09 Aug 2026 11:55:44 +0530

One of my goals this year is to learn to dance.

I haven’t done much about it, but I did snatch one opportunity. At VizChitra, Ashok Kumar led an Afternoon Rhythm “where the drum sets a beat and you find your place within it.” He invited volunteers on stage.

I usually volunteer for uncomfortable things (a habit from school days), so I briskly walked to the stage, waited for a few others to join, then started dancing to the beat.

I’ve been told (as a kid) that I was graceful. I haven’t danced in decades. I tried a few tentative moves.

Then my mind voice went, “Screw this. Just f***ing dance. Don’t worry about looking good.”

I closed my eyes. I danced to the beat.

Here's the video

This is the part where I’m supposed to say something poetic about how I felt. But it wasn’t like that.

Some parts looked awkward. Felt awkward.

Some parts flowed. At least to me.

But for a few minutes, I was back at hostel, dancing to Ilayaraja’s beats in my room, all by myself.

One of my goals this year is to learn to dance. Even more so, now.

A Tea Perspective

Sun, 09 Aug 2026 11:49:12 +0530

When I was at school, I assumed that teachers in the staff room would mostly be discussing students, teaching, how to improve things, etc.

A few years after I graduated, I spent time with my teachers in the staff room, and realized that the conversations are far more mundane. The main topic of discussion (which went on for what felt like half-an-hour) was: Why was the tea at the high school staff room far inferior to the one for the junior school staff room? What could they do about it?

It was a bit of a shock for two reasons.

What!? Teachers don’t spend every second of their time thinking about students and teaching?
Tea!? Who cares about tea? Why care about tea?

That day, my perspective changed.

25 years later, my perspective changed again. I wasn’t a tea drinker. I started a year ago when I switched my daily lunch to Ya Kun Kaya Toast’s Set B: Kaya peanut butter toast (which is heavenly) + eggs + tea. I always get it from the same branch across my office.

There’s one guy who I think makes it slightly sweeter, and a lady whose tea is more bitter. I spend a fair bit of time every morning (and sometimes it feels like half-an-hour) agonizing over who will be making my tea today.

Never thought the day would come, but this is one of my highlights of the day: having his tea for lunch.

Comic art style prompts

Sun, 09 Aug 2026 11:09:26 +0530

Many people commented that they liked my comic illustrations and asked how I create them. Here is my process:

Paste a reusable prompt fragment that’ll take any content, think about what to draw, then draw it.
Paste a style variation for different comic styles (optional).
Paste the content itself and run it.

I use ChatGPT with gpt-image-2 more often than Gemini with Nano banana 2.

Here are the prompts.

STEP 1: Reusable Prompt Fragment: I have a few of these right now:

Comic page prompt fragment:

Draw this as a full-color explainer comic page (portrait) - sequential explanation, friendly narrator, diagrams embedded inside panels, visual metaphors, self-aware captions, and clear cause-and-effect storytelling.
Style: expressive characters, comic-style ALL CAPS, vibrant modern colors, clear visual hierarchy.
Prefer pictures over words. Use recurring visual metaphors so the reader understands the idea even while skimming.

First, write a memorable storyline that captures the most important points to convey - as a single cohesive story.
Just reading the storyline should communicate the entire message unambiguously.
Critique the storyline: what is confusing, doesn't flow, or has low impact? Revise. Repeat until the storyline is GOOD!

Draw each storyline element (typically a sentence, but sometimes a continued phrase, or multiple sentences) as a panel's caption. (If there are 8 panels, there must be 8 storyline elements)
Each panel's image should support and strengthen its caption - and reinforcing past panels / anticipating future panels where helpful.

Example:

Comic strip prompt fragment:

Draw this as a simple black and white line drawing comic strip (1:1) with minimal shading.
Single panel.
Style: expressive characters, comic-style ALL CAPS.
Prefer pictures over words.
No need to cover everything - just one key item is enough - e.g. the funniest, most important, or most surprising point.
Convey the INTENT of the point. An apt analogy that visually communicates instantly might work better than a literal depiction.
Keep it funny. The strip itself should make readers laugh.

Example:

STEP 2: Style Variation: This is optional. Here are examples of a few different styles:

These are cataloged in my LLM Art Style gallery (see blog post).

STEP 3: Paste whatever content I want to illustrate. Some examples are:

Transcripts of my talk
Contents of my blog post
An email reply I’m sending

The main insight is that ChatGPT and Gemini can think about what best to draw, and then draw it. So I can, with some careful prompting, delegate the comic design to them for any content, making this an automatable flow.

Things I Learned - 09 Aug 2026

Sun, 09 Aug 2026 00:00:00 +0000

This week, I learned:

Kamakoti: “Entry (to the course) is relatively easy but the exit is extremely hard”. Generalizing, quality is determined by the exit criteria; loosening entry criteria is just openness / diversity. The Hindu
“Once I have a persistent system that I pay to keep thinking, learning, and acting 24/7, I think that will decisively look like AGI.” - Dan Shipper
AI has expert-level capabilities in many (increasing) tasks #ForNow. If your edge is OUTSIDE of those, use AI for other tasks you couldn’t do before, insourcing or expanding horizontally. But your edge may be short-term - so move upstream / specialize. Your competitors’ edge may be short-term, too - so plan to attack.
Analyzing Anthropic Economic Survey, it looks like people in rich countries are asking Claude for advice (explain this spreadsheet) while poor countries are asking Claude for output (build this website) #ForNow. Maybe because rich users already have tools / people that create output for them?
Humans can’t define all laws of language but LLMs have learnt them anyway. What if there are laws of nature that humans can’t understand but AI can? Actually, this is already true of black-box models (loan approvals, weather forecasts, …) where benefit/control > understanding. But as data & compute scales, this might “solve” entire fields like psychology, economics, etc. Noah Smith
In each area, there might be a limit to how much intelligence is possible/useful. For example, we’re pretty good at recognizing food and emotions - there’s not much benefit / possibility of more intelligence. But we can copy and share this intelligence - and that might help more than we think. Noah Smith
The ChatGPT Dropbox plugin can read Markdown files if you specify the path, but can only read PDF, Word, PPTX, Excel, etc. when searching. It cannot update files on Dropbox, but can add and delete. #ForNow
Given how long agents run without mistakes, verification is increasingly “drift correction”. You can’t spot it easily. Learn writing specs that EXPOSE drift. Build and test against “oracles” (verification systems). Reduce cost of error.
Permissions, in the context of multiple agents, is complex. If agent A can read my email but wants to consult agent B, can B see the email? We’d need to make permissions pretty specific, like:
- principal: “anand”
- agent: “agent-17”
- purpose: “insurance-coverage-check”
- allowed_data: [“email:read”, “dropbox/notes:read”]
- allowed_effects: [“email:send”]
- audience: [“anand”]
- expires_at: “…”
- delegation_depth: 1
After struggling to understand where to apply loop engineering, here’s my guess. If you have a metric (or something really well defined) that you want to optimize, and a single agent iteration isn’t enough, loops are a way to get there. Kaggle competitions, benchmark optimizations, etc. are examples. This means that any complex system that you can benchmark (or at least where you can robustly compare results) is loop engineerable. (This means that the ability to benchmark, and using agents to benchmark, will become a key ability.)
Ontologies, state machines, etc. can be used to create verifiable systems, e.g. nodes become states, relations are valid operations. That’s great for building verifiable systems (leading to things like LEAN). Of course, a key skill will be knowing what to put into the state, what relations to allow/disallow, what reflects reality well, how it might evolve (e.g. temporal graphs), how that might change in the future, etc.
- Having said that, this is just creating a neural network of sorts - so according to the bitter lesson, we should just toss data at an agent and have it build a graph (or not) as required. BTW, I shared this with a bunch of speakers at Data Hack Summit who were speaking about knowledge graphs. There was silence for a while. Then, gently, they all agreed.
Some people blab. Interrupting with a question is a good diversion mechanism. Some blab even after that. Exiting politely is both wise and surprisingly un-rude.
To control your mental state, breathe slowly. 5–6 times/min for five minutes (that’s longer than I thought was needed), exhaling slower than you inhale. PubMed
Once a ChatGPT conversation uses a developer plugin, #ForNow it refuses to use other plugins. So, if you need to use a GMail plugin AND a plugin you built yourself, you’d need to use the GMail plugin first, get stuff into the chat, then switch over to yours. I suspect conversations with developer plugins might not be accessible when using other plugins, too - but that’s untested.
There’s a jagged edge of AI adoption as well, not just AI capability. Several organizations limit users to weaker agents #ForNow (e.g. only Microsoft Copilot or Gemini). Many have never seen the power of Codex or Claude Code on their systems. It’s hard to convince them that AI can do much more than they think.
There’s a “data engineering” industry incentivized by structuring data. This is partly enabled by poor enterprise agent adoption #ForNow (e.g. Microsoft Copilot). The sequence works like this: “AI does not solve something with the data it’s given. Let’s structure the data. It solves it. Therefore, we need to structure data - all data.” The alternative which I believe is: agents will structure it themselves.
I noticed that when you submit a prompt on ChatGPT, it changes the URL to https://chatgpt.com/c/WEB:... and once it starts processing it on the server, changes it to https://chatgpt.com/c/... giving it the actual conversation. So, if you see a WEB: in the URL #ForNow, make sure you copy the prompt before reloading the page - because it hasn’t been saved or sent to the server.
I assumed inflammation was mostly a bio/chemical process. Looks like neural signals are involved, too, and electrical simulation can control inflammation. This leads us to a new territory: bio-electrical medicine.
The Anthropic Economic Index indicates that, on average, if you prompt Claude like an 8th grader, it responds for a 9th grader. Does that mean (a) that more sophisticated prompts get a better response, and (b) if you repeatedly meta-prompt, you increase the sophistication by about a year each iteration, and hence can get very smart prompts by just getting out of the way and with little hope of understanding the question? This might actually make sense if AI will action the result without you needing to understand.
The geometric mean is always less than or equal to the arithmetic mean. This is why a “smooth” 8% return is worth much more than a “wild” 8% return. @lumenxbt
Quantum cryptography can give us unclonable encryption, i.e. if someone copies a message midway (or you publish it), you can’t independently decrypt both. We knew how to do this in 2020. Now, ChatGPT helped “indistinguishable security”. Between 2 messages, people can’t figure out (e.g. from the length, or other attributes) which message is which. Gemini
Agents can record network requests into a HAR file and reverse-engineer an API for many websites. More efficient than browser control. dax
The Anthropic Economic Index dataset is on Hugging Face - released quarterly #ForNow. The longitudinal analysis is likely to be interesting.
BusinessCaseBench solved over 238 business cases with AI agents and they’re doing well and improving #ForNow. Not surprising. Frontier AI performance across the business disciplines
OpenAI Presence shows a pathway for deploying agents. Deploy for a specific job, with only required access to knowledge and systems, company defined policies for approval, agent periodically reviews logs & escalations and proposes updates for testing and approval.
A lot of work people are doing on ChatGPT is OUTSIDE their area of work. “… a substantial part of work-related ChatGPT use is from users expanding their role.” OpenAI

Questions I was asked

Week ending 09 Aug 2026

Question: Is this (Email AI with LocalMCP plugin) just search on steroids?
Answer: Yes. Search on steroids is not a bad mental model. The deeper capability is the ability to loop like crazy - keep hitting a problem with tools until it actually gets solved.
Question: How do you build enough trust to let agents take multiple steps without constant approval?
Answer: Don’t try to convert everyone. Leave the early adopters alone; show the middle working examples and let them try carefully. When the middle moves, the rest will catch up themselves.
Question: If AI accelerators become outdated in two weeks, how do you keep up?
Answer: Build accelerators for accelerators. Instead of investing in a benchmark, build a benchmark-builder from production logs; leapfrog one step and plan for agents to automate today’s specification and verification work.
Question: How do we govern all the AI apps employees are creating?
Answer: First ask, “Does this need governing?” Personal use: do whatever you want. Shared apps: review them. Don’t say, “You cannot do X unless it is governed.”
Question: Where do you stand on “code with AI, code without review”?
Answer: Usually review it, with AI helping find problems. But for throwaway code, or code agents write for themselves, or incidental to a business output you can verify, validate the outcome instead.

Mistakes I made

Week ending 09 Aug 2026

I described the MoSPI telecom results as applying to youth aged 14–24 and referred to Daman and Diu as the geography.
Correction: The CMS: Telecom 2025 tables use the age group 15–24, not 14–24. The official State/UT geography is Dadra & Nagar Haveli and Daman & Diu. When quoting striking subgroup percentages, I should preserve the source’s exact denominator and geography. Evidence: MoSPI — NSS Report No. 593: CMS Telecom, 2025
LOW · FALSE
I said agents “don’t hallucinate, not anymore.”
Correction: Hallucinations have fallen substantially in newer models, especially when they can search, use tools and verify, but they have not disappeared. OpenAI still explicitly evaluates hallucination rates and warns that ChatGPT can confidently produce incorrect or misleading outputs. Evidence: OpenAI — Does ChatGPT tell the truth? OpenAI — GPT-5.6 System Card
HIGH · OVERSTATED
I said “GPT-4o-mini is both better and cheaper” than GPT-4o and suggested simply switching.
Correction: GPT-4o mini was much cheaper, but it was not simply a more capable GPT-4o. Model quality depends on the task. OpenAI’s own evaluations, for example, showed GPT-4o substantially more accurate and less hallucinatory than GPT-4o mini on SimpleQA and PersonQA. The right advice is to switch an old expensive model only after benchmarking the cheaper candidate on the actual workload. Evidence: OpenAI — o1 System Card, Table 3
MEDIUM · OVERSTATED