S Anand

My browser is my notebook and task list

Tue, 10 Nov 2026 15:10:47 +0800

Recently, I’ve started using my browser both as the primary workspace (slightly more than VS Code) as well as my TODO list. That’s mainly due to better tab management and better tools.

Tab Management

The first big shift was turning on Vertical Tabs on Edge. Now, I can see and manage far more tabs. Typically, I have 80-200 tabs open now.

The next big shift was using tab groups. With vertical tabs, this makes tabs like a folder.

Since I was accumulating too many tabs, I needed a management system. I decided to apply the same principle as my TODO lists:

All current work goes into a “WIP” tab, or task-specific category like Read, Lego, TDS, etc.
All previous week’s WIP tabs are renamed “WIP YYYY-MM-DD” for that week.

I also created a “Tab Notes” extension to help me make notes on each tab, e.g. what I should do next, when I opened it, etc.

… and updated my “Anand Chrome Tools” extension that lets me change the title of a tab, telling me what I should do next.

Over time, I didn’t really need the extensions - the tab groups work well enough.

I also use an edge CLI that lets me export / import all my tabs. I run this daily to backup my open tabs. So, even if I close several tabs accidentally, I can restore most of them from yesterday.

Of course, a big unlock is that I use browser tabs as slides almost exclusively these days.

Tools

I find four capabilities really effective when extending browser capabilities.

Bookmarklets let me change any site in almost any way.
Extensions let me do more - even changing the browser itself.
Chrome Devtools Protocol (CDP) lets me write external scripts that control the browser.
Browser profile data lets me write programs to access browser history, cookies, etc. even without controlling the browser.

Bookmarklets are a big part of how I extend my browser. I have a pretty extensive list of bookmarklets.

Some of the most used ones are:

AI Scrapers - where I copy ChatGPT, Claude, or Gemini chats / prompts to save or pass to another agent.
WhatsApp Scraper lets me copy an entire WhatsApp thread as Markdown or JSON.
Meet Captions lets me save captions from Google Meet or Microsoft Teams. I can then ask AI “What should I say next?” and pass it the WIP file as context.
GMail header lets me copy the mail header to pass to an agent, e.g. “Reply to this email…”
Dark Mode enables dark mode for GMial and Hacker News, which don’t have this natively and I use a lot.

To make it easier to access these bookmarklets, I set up “Anand Chrome Tools” extension so that Ctrl + Shift + Space acts like the command palette in VS Code. I can pick a bookmarklet to run, a tab to switch to, or a custom command (like setting titles).

I also summarize what I browsed each data as part of my activities update - which gives agents a bit more context about when and why I browsed a tab.

It feels like I’m just scratching the surface of what’s possible, actually. Now that agents can write most code I can imagine, imagination may be all we need.

Linkedin List Feeds

Tue, 10 Nov 2026 13:17:24 +0800

X has lists. You can follow posts by any list of people - and share the list.

LinkedIn has this feature too. It’s tougher to use, but easy to share.

Visit the posts search page
Select the “From member” filter and add members.
Bookmark or share the link.

Then I asked ChatGPT to recommend people to follow. Here’s my current list:

List	People
AI	Rama Ramakrishnan · Arvind Narayanan · Simon Willison · Rohit Krishnan · Sayash Kapoor · Benedict Evans · Seth Lazar
FOSS	Siddharta Govindaraj · Sai Rahul Poruri · Kailash Nadh · Daniel Stenberg · Daniel Lemire · Phil Eaton
Edu	Greg Wilson · Bonni Stachowiak · Dan Meyer · Robert Talbert · Marc Watkins · Justin Reich
Data Viz	Curran Kelleher · Rasagy Sharma · Nadieh Bremer · Dan Saffer · Alberto Cairo · Amanda Makulec · Santiago Ortiz · Georgios Karamanis · John Burn-Murdoch
Data Viz Orgs	Our World in Data · Data For India · Datawrapper · The Pudding

This is the combined list for all 28 people

Converting TUS Nesstar Files

Mon, 09 Nov 2026 16:00:12 +0800

The Time Use Survey 2024 data released last year lets you download a ZIP containing a 900 MB .Nesstar file.

(This is after you register - but the registration is free, and though it requires a reason, it’s automated.)

Nesstar is Norwegian software that’s now end of life. MOSPI provides this guide - with distorted screenshots - explaining how to install the software on Windows.

But I use Linux. The software is end-of-life. Gaah!

However, I have ChatGPT.

I uploaded the .Nesstar file and prompted ChatGPT to “Convert this TUS data to an Excel and Parquet file format so I can download it.” The first attempt failed (the conversation became too long) but the second attempt let me download the data as .parquet and gave me a thrift-based uv script to convert the file to .parquet. (nesstar-reader might work, too.)

The ease of this converstion opens up a wide variety of possibilities in my mind.

Accessible data. Re-publish data in archaic formats. For example, Saket recently created nesstarR and maintains mospiR - making any MOSPI dataset more accessible.
Retrospective research. Re-run experiments from old raw data. For example, NASA found two moons from old Voyager data. This is a good reason to capture more data in raw formats - and also not worry too much if we can’t, since future intelligence might probably overcome limitations.
Rewrite regulatory software. For example, OpenTax is an agent-engineered US federal tax calculator. OpenBooks aims for EU-compliance. It would be nice to replace Singapore’s BizFinx, Pharma Open eCTD publishers, and build an open regulatory compliance infrastructure.

Things I Learned - 08 Nov 2026

Sun, 08 Nov 2026 00:00:00 +0000

This week, I learned:

Facemorph.me implements face morphing via StyleGAN2 (among the best face morphing implementations I’ve seen) and exposes it via a free public site #ForNow.
My reading from The Sequoia of Tomorrow:
- Money’s in the Bay Area. People, move there.
- Money’s in a few big winners. Maybe they can find them easy and don’t need to scout?
- Money’s in speed and fewer deciders. Partners don’t need to wait for and convince each other every Monday. Risky, but faster.
Looks like you can use Codex via an API, since the OpenAI Agents 0.22.0-alpha.1 release is available. “This alpha introduces the embedded Codex harness in the OpenAI Agents SDK for JavaScript and TypeScript.” Via Daniel
The amount I’m delegating to agents is small but growing. In Aug-Sep, a median session took Codex 3.3 minutes and only 4% of my prompts took 20+ minutes. In October, the median session took 7.5 minutes and 13% of my prompts ran 20+ minutes.
ChatGPT image generation will be a separate quota from Codex #ForNow.
Armin used Qwen3-TTS to detect voices. I ran a quick benchmark to see how well it did on my transcripts. Answer: Not very good. 32% precision, 84% recall.
“Professors nowadays seem to have become professional fundraisers who do a little research on the side.” Also, “Turning the other cheek turns out to have selfish advantages.” Paul Graham
Looks like we’ve learnt how to get a neuron to fire by shining a light on it. It’s called Optogenetics.
Voice remixing is now available in Gemini 3.8 Flash and it did a pretty good job of replicating my voice as well as creating new voices based on a prompt. via X. Qwen3-TTS supports custom voices too and can run on your laptop with just 1.7B parameters.
This (interestingly summarized) Machines of Loving Grace suggests factors that bottleneck intelligence (like in clinical trials, ) - and hence become the focus:
1. People move slowly or impose constraints.
2. Data isn’t available.
3. Nature (biology, chemistry, engineering) works only so far.
Nathan Baschez sees agents factories as the next step beyond “All employees use agents”. These are shared workflows maintained centrally that the organization can get work done with. That packages context (skills, prompts, docs, data) along with environment (tools, permissions). The main advantages are centralization (few experts/owners enable everyone) and execution (acting in the real world).
WACZ seems to be the modern way to archive websites. Unlike MHTML, it can handle multiple pages.
In the last 10 months (Jan-Oct 2026), I prompted Claude, Claude Code, ChatGPT, and Codex about 12,000 times. On average, I talk to AI ~40 times a day. That’s 80% of my “Talk to AI 50 times a day” recommendation. I still have some way to go.
Phone spam marketing has reached an elevated level of obnoxiousness. I get calls from automated recordings often. But today, Airtel Black call me and put me on hold for two minutes. Wow! Then the music stopped and the line just hung. So, they put me on hold and ghosted me. And the reason I picked up the call was because their Internet was down - so a triple whammy!
Not only have I started enjoying tea a lot, I quite look forward to it. Even to the point where a high point of my day is when someone serves me tea, unasked. As my daughter commented, I’m getting old.
wacli is a CLI interface for WhatsApp. Unofficial, and works well #ForNow.
There’s no clear, good, ChatGPT search / export CLI that I am comfortable with #ForNow. ChatGPT

Mistakes I made

Week ending 08 Nov 2026

I said “If they do it better, then TDS has taught them that skill” while proposing to measure learning by comparing students’ performance on successive questions.
Correction: Better performance on a later question is evidence of improvement, not proof that the course caused it. Differences in question difficulty, prior knowledge, independent practice, agent assistance, and other factors may explain the change. Establish comparable assessments and use a suitable comparison group or research design before attributing improvement to teaching. Institute of Education Sciences: Research Design Standards.
HIGH · OVERSTATED
I said “the accelerometer tells me how much the phone is moved” while explaining why measuring distances and reconstructing scenes with a phone should be simple.
Correction: An accelerometer measures acceleration, not displacement or distance. Estimating motion by integrating acceleration accumulates errors. Reliable mobile augmented reality combines camera observations with inertial measurements and may use depth sensors; distance estimates require suitable visual geometry and calibration. Android: Motion sensors; Google ARCore: Fundamental concepts.
MEDIUM · FALSE
I said “map the approximate meaning of individual neurons” and “change the weightage” of a feature using Gemma Scope while discussing research on controlling LLM behavior.
Correction: Mechanistic interpretability tools such as Gemma Scope use sparse autoencoders to identify features represented across model activations, not necessarily individual neurons. Individual neurons can encode multiple unrelated features. Feature steering can intervene on activations during inference without changing the model’s trained weights. Such interventions can influence behavior, but identifying a feature does not establish a simple, reliable control mechanism. Google DeepMind: Gemma Scope · Neuronpedia: Feature steering
MEDIUM · OVERSTATED
I said “Rambler … [is] not available via GBoard” in a weekly learning digest about Android voice typing.
Correction: Google’s Gemini-powered Rambler voice input is built into Gboard on supported Pixel 11-series devices. It requires an updated Gboard and supports limited offline functionality. A separate keyboard application is not required to access Rambler on supported devices. Google: Rambler voice input on Gboard.
MEDIUM · FALSE
I described Clairic as “a Decision model (or System One model)” in a technical learning digest.
Correction: Cloudflare classifies Clairic as a System One extraction model, distinct from decision models such as Clef. Decision models select among predefined choices; Clairic extracts, computes, estimates, or predicts typed values from inputs. Both belong to the broader System One model family. Cloudflare: Introducing Clairic.
LOW · FALSE
I included “1923: White chocolate” in an AI-assisted timeline of chocolate innovations.
Correction: White chocolate emerged commercially in the 1930s, not 1923. Nestlé dates the launch of Milkybar to 1936 and records the introduction of Galak white chocolate in approximately the same period. Nestlé: Milkybar history; Nestlé company history.
LOW · FALSE
I included “1923: Filled chocolate bars” in an AI-assisted chronology of confectionery inventions.
Correction: Filled chocolate confectionery existed decades before 1923. Fry’s produced chocolate-coated Cream Sticks in 1853 and launched its mass-produced, fondant-filled Chocolate Cream bar in 1866. The Chocolate Museum: Fry’s history; Oxford Companion to Sugar and Sweets: Filled chocolates.
LOW · FALSE
I included “1971: Chocolate eggs (Cadbury Creme Egg)” in a chronology of chocolate innovations.
Correction: Chocolate Easter eggs predate 1971: Fry’s produced hollow chocolate eggs in 1873, and Cadbury introduced its first chocolate Easter egg in 1875. The year 1971 marks the launch of the specific Cadbury Creme Egg product, not the invention of chocolate eggs. UK Companies House: History of chocolate eggs; Cadbury: Company timeline.
LOW · OVERSTATED

Questions I was asked

Week ending 08 Nov 2026

Question: How do I get useful feedback on a visualization when the decision-maker has no time? Asked while discussing how to test dashboards intended for busy managers rather than fellow students.
Answer: People refuse feedback because it takes too long or they don’t see the point. Make the task tiny: show two visualizations for 10 seconds, ask which communicates better, then ask why.
Question: How will collaboration work when an adaptive exam gives every student different questions? Asked while designing an exam whose difficulty and next question depend on individual performance.
Answer: Weaken collaboration, but don’t destroy it. Add real-time challenges where several students enter a virtual room and must accomplish something together. They can bring agents, but coordinating with other humans becomes essential.
Question: Are we measuring what the student learned or merely what AI helped the student accomplish? Asked while planning to compare students’ performance on repeated skills across assignments.
Answer: Be precise about what learning we claim. The unit of learning may be the student plus their agent, context files, tools, and social network. If someone builds resources that make them consistently more capable, that may be a valid learning outcome even when the knowledge isn’t entirely in their head.
Question: How do we measure whether students have learned to combine skills, rather than merely solve individual questions? Asked while reviewing longitudinal exam results for Docker, APIs, and other technical skills.
Answer: Mine existing assignments for questions combining three or four earlier skills, then compare the same students’ performance. Better still, deliberately design later questions to measure transfer. Don’t just test whether they learned Docker; test whether they can use Docker, APIs, and other skills together.
Question: How do we explain what students are learning when an AI-native course has no fixed syllabus? Asked after students complained that they couldn’t predict the questions or describe what they’d learn.
Answer: If I can structurally, deterministically train you in something, why wouldn’t I train a model instead? Most courses teach known skills that agents may eat up. This course explores where you can create value that agents cannot yet replace.
Question: Can we use AI instead of humans to evaluate whether a visualization communicates correctly? Asked while comparing ways to test visualizations, including simulated audiences and adversarial interpretation.
Answer: No, not as a complete substitute. Agents perceive things differently. Humans unanimously recognized a drawing as a donkey while models guessed cow, sheep, or horse; the reverse happened with an owl. Use AI to generate failure cases, independently recalculate data, and suggest alternative designs. Still test with humans.
Question: How do we teach communication skills when the person who builds something is often not the person who presents it? Asked while discussing students who produce excellent projects but struggle to explain or sell them.
Answer: I’m experimenting with letting students publish services and sell them to each other for credits—anything from data cleaning to helping with an exam. They must find customers and negotiate. The design needs work: some students trade credits without actually delivering anything.
Question: How do you preserve human interaction when AI makes you increasingly self-sufficient? Asked after observing that working extensively with Claude had made him more isolated from colleagues.
Answer: I have the same problem. Conversation is an agent-era-proof skill, so I use agents to strengthen it. Every week, I ask one to review my conversations, identify people I should thank and why, and then I send those messages.
Question: How can multiple agents work together on the same task? Asked after seeing an AI-agent swarm collaborate and comparing it with his single Claude assistant.
Answer: Start with one agent running hourly: read pending.txt, finish whatever it can, add new work, and mark anything blocked. Put another agent on a different machine and give them a shared folder or request queue with clear boundaries. Let them exchange work directly. Remove the human relay and you’ve got a swarm.
Question: How do people build agents that book flights and do things without being prompted each time? Asked after demonstrating several AI-built internal applications and hearing about personal agents.
Answer: Telling Claude to book a flight is just using Claude. The interesting part is what it does unprompted. Schedule it to scan your context daily: what have I promised but not delivered? What commitments from others are pending? What possibilities am I missing? That’s a way of expanding ambition, not merely automating tasks.
Question: Can AI reliably follow our existing presentation design, or should we let it create its own? Asked while reviewing AI-generated investment-banking presentations that still needed manual design work.
Answer: Separate instruction-following from design. I’d trust ChatGPT more to preserve a supplied template while changing only the words, and Claude more to invent a good chart without a sample. Don’t insist on one agent: have one produce the deck and another interrogate it for mistakes.
Question: How can I trust a RAG benchmark when the LLM itself generates the test questions? Asked while comparing reranking, agentic RAG, and knowledge graphs on financial filings.
Answer: Start with the audience and the decision they need to make. An IT director should immediately recognize the questions as representative. Classify them by what the infrastructure must do—lookup, comparison, searching across documents, inference—and test roughly six questions per pattern. Keep the questions short enough to understand at a glance.
Question: How do we design an AI-permitted exam question that a shared solver won’t make trivial for everyone? Asked while proposing questions that interrupt an agent’s reasoning or alter instructions embedded in attached files.
Answer: Changing just a number or phrase won’t work; one student can build a solver and share it. Personalize the actual reasoning problem, not merely its parameters. I want many students to learn how to use agents cleverly—not one student solving everything, and not a question so hard that almost nobody learns.
Question: If agents can build software so cheaply, what does a product manager do? Asked after discussing a custom drawing app built in minutes and a conventional product pilot that took nine months.
Answer: Maybe there’s only one software now: agents. The product manager discovers users’ needs and workflows, gives agents the right context and tools, and sets up the feedback system. Log everything, read the logs, decide what to change, and optimize. Building that data-driven decision system may be more valuable than building the application.
Question: What can programmers learn from managers about the constant context switching caused by running many agents? Asked after describing how checking idle agents and parallel tasks was destroying his flow state.
Answer: Management is a useful analogy. Learn how to specify, verify, delegate, and manage an intelligence that may know more than you. Also study patent clerks, judges, auditors, and detectives: they’ve developed ways to make decisions despite incomplete information and experts knowing more than they do.
Question: Who will pay for software when anyone can build their own with AI? Asked while discussing inexpensive personal software, liquid software, and the declining scarcity of programming.
Answer: Software may become liquid: people modify and fork applications for their own needs. Things consumed by agents can increasingly be packaged and optimized as products. Services may tilt towards the performing arts—not literally, but towards human performance sold to a human audience. The commercial models are still unclear.

Picture stories we understand and like

Sat, 07 Nov 2026 21:41:08 +0800

Students of my course were given an assignment.

Submit a picture story (no words) that others will understand and like.
Explain others’ picture stories and rank them.

Here are the results.

Firstly, why this exercise? Because it teaches three skills I believe are important in the AI era:

Creating unique content
Communicating with clarity
Marketing your ideas

This is exactly what they were evaluated on.

Uniqueness: How similar was their story to others'?
Understanding: How well did others understand their story?
Preference: How much did others prefer their story over others'?

Here are some stories I enjoyed. Click for details.

One interesting relationship is uniqueness and preference. The two are negatively correlated. I’m not sure why.

The least unique stories were about these three patterns - and there were so many of them!

If people can’t come up with original stories easily, I guess there’s a concrete skill to learn and a way to differentiate capability.

The most unique stories were reasonably interesting. They included:

The top scoring story was fairly unique, too: The Bhagavad Gita.

I was surprised at how interesting some of the low scoring stories were. For example, I thought this photographer deletes a baby panda photo to protect it was really powerful.

It ranked low on understanding (3 people completely mis-understood the story but since they labelled the story “A thirsty crow” or “A steam engine…” I think these were errors or junk). It ranked only ranked mid-way on preference.

So, maybe there’s more to learn than what I was measuring.

Still, it was good to know that we can teach and evaluate useful skills even when students can use AI and collaborate.

Toenails

Sat, 07 Nov 2026 18:25:35 +0800

Note: This continues my nasty habits series: licking, scraping, and nibbling - in which I mentioned fiddling with nails.

There are three irreplacable possessions in my room:

My phone, because it authorizes lots of services.
My laptops, because they have un-backed-up data.
My toenails, because the good ones are rare.

Here are two good ones I’ve had for a few months. The ones best for fiddling are from the toe thumb. They typically have:

A large, thick, smooth surface that I can run my finger over
A thin, tapering edge that’s springy and I can twang it
A ragged edge that I need to be careful of
Ridges and grooves that I can pluck at
An irregular bend I can pull at to straighten
A sharp tip I can poke with
… and so many other delights.

Seriously. These are irreplaceable. How often do you get a toenail with all of these features? (Once a year, if I"m lucky.) That makes them more irreplaceble than my phone or laptop.

The biggest threat to my possessions are our maids (we get different ones periodically). Apart from turning off my remote laptop, they also sometimes discover my secret stash of toenails and throw them away. The last time this happened, it took me an entire quarter to get a moderately satisfactory replacement.

Unlike the laptop shutdown, this is harder to prevent. It would mean telling my wife about the toenail stash, which could lead to a permanent ban.

Using Real Faces in Comic Stories

Sat, 07 Nov 2026 16:40:52 +0800

I was reading Asterix on my flight from Brussels to Dubai when I remembered a colleague’s request. I typically convert my workshops into narratives like this and this. They wanted another event converted.

Usually, I have the full transcript of my events. In this case, I just had a few paragraphs of notes. Far from enough.

On the other hand, I was reading Asterix. And AI’s my magic potion. So I figured:

Why not create a comic story?
Using real faces?
While reading Asterix?

Step 1: Story to prompts

So, step 1. Comic story. That’s easy. I’ve researched multiple comic styles and have a sense of the style that might suit. So, I told Claude Opus 5.5 High, roughly:

Read [the email]. Let’s use ChatGPT to create a black-and-white comic strip series narrating what happened.
Sort of like Calvin and Hobbes or Peanuts or even Scott McCloud’s Understanding Comics.
Maybe a dozen strips of 2-4 panels each? Or 4 pages of 8-14 panels each?

Think like an expert storyteller and comic artist.
Research ChatGPT’s latest image model - good and bad.
Give me a series of prompts for ChatGPT.

I want people to look like their photos. Create prompts.

That gave me a nice story using the black cat in the Black Cat Bar as the narrator.

Step 2: Get profile photos

I asked ChatGPT with GPT 6 Sol - High to search for the people:

Go through this email thread.
Download a few photos (preferably profile pictures) of each of the people mentioned in the email.

… and corrected it:

You have Namit’s picture instead of Ankor’s. Also, photo 2 isn’t what’s in the source you indicated.
These worry me a lot.

Go through every single one of the photos you’ve identified and check if it’s really from the source and the images match.
Make sure these ARE, in fact, the people that attended the event - and flag any uncertainty.

This worked well. There weren’t any mistakes, but there were a few people and we weren’t sure if THEY were the ones in the event.

Mid-flight, I saw that they’d sent me actual attendee photos from the event. So:

Match the profile pictures against the attendee photos.
Label who is present where in which picture.
Identify unknown people - and see if you can search for them online,
e.g. to see if they match any of the people we couldn’t find images for.

That gave me an excellent base.

Step 3: Create a style guide

This was a clever step by Claude. It told me to create a photo style guide in four steps. Click on each to expand the prompts.

1. Create a character model sheet for the cat

STYLE BIBLE — “THE CONVERGENCE ROOM” (a 13-page black-and-white comic)

LOOK

Black ink on pure white paper. No colour, no grey wash, no sepia, no paper texture, no gradients.

Shading only with solid spot blacks, fine cross-hatching and sparing dot screentone.

Classic newspaper comic-strip craft: confident brush-and-ink line with thick-to-thin weight, loose and hand-drawn, slightly imperfect. Not vector-clean, not manga, not photoreal, not 3D.

People: warm, simplified caricature with normal adult proportions. Likeness comes from a few signature features (face shape, hairline, glasses, facial hair, build), not detailed rendering.

Characters are told apart by silhouette and by black/white value of their clothes, not colour.

LAYOUT

Square 1:1 image, highest resolution available.

Hand-ruled black panel borders, even white gutters. 2×2 grid unless the script says otherwise.

Hand-lettered title bar across the top: “THE CONVERGENCE ROOM • [NO.] • [TITLE]”.

Read left-to-right, top-to-bottom. Vary shots: wide establishing, medium, close-up.

Flashback panels have wavy borders.

LETTERING

Hand-lettered ALL-CAPS comic lettering, large and very legible.

Dialogue in white oval balloons with tails pointing at the speaker.

The narrator, TOMBSTONE, never has balloons, only rectangular caption boxes (white fill, black border).

Speaker tags such as “— A DIGITAL BANK” are small caps tucked under the balloon.

Render every piece of text EXACTLY as written in the script, letter for letter. Add no other words. No logos or brand names anywhere. Screens, signs, papers and badges are blank or scribbled unless the script gives their text.

WORLD

The University Club of San Francisco on Nob Hill: red-brick Italian Renaissance clubhouse, arched windows, deep dentil cornice, chimneys; cable cars on steep streets.

Dining room: dark beamed wooden ceiling, brass candle chandelier, tall windows with wooden louvred shutters and city lights at night, candle wall sconces, round tables with black tablecloths, wine glasses, carved high-backed wooden chairs, patterned carpet.

The Black Cat Bar: curved wooden bar with a brass foot rail, framed pictures on deep-toned walls, small brass lamps.

Detailed backgrounds in establishing panels; minimal backgrounds (a few lines or white space) in dialogue panels.

CHARACTERS

TOMBSTONE (narrator): a plump, loaf-shaped black marble cat statue on the bar. Smooth rounded forms like a river stone, small rounded ears, solid black with one thin white highlight curve along its back, long tail curled around its body. Eyes: two white almond shapes with black slit pupils; dry, deadpan expression. When “alive”, only eyes, ears and tail move, except when the script says it walks. Match the TOMBSTONE in the attached references exactly.

HOSTS (recognisable): ANKOR, RONNIE, NAMIT, SUKRUTH. Match the attached CAST SHEET exactly; same face, hair and costume every time.

GUESTS (anonymous): smart evening business wear. Invented faces that do NOT resemble the hosts, any reference photo or any real person. Vary age, gender, ethnicity and build. Identify them only by the role tag in the script.

Image 1 is a photo of the real Black Cat Bar. The dark, chunky, rounded cat statue on the left end of the counter inspires our narrator, TOMBSTONE. Use the photo only for the cat’s plump loaf shape and the bar. Do not copy any artwork on the walls, and add no people.

Draw a character model sheet for TOMBSTONE on a white background, square, in the STYLE BIBLE’s black-and-white brush-and-ink style. Seven small drawings, evenly spaced:

Sitting as a still statue on the curved wooden bar, eyes closed.

Same pose, eyes open, deadpan.

Eyes narrowed, suspicious.

Eyes wide, startled.

Alive, walking across a black tablecloth between wine glasses.

Looking straight out at the reader, one eyebrow-ish ear tilted.

Pure black silhouette. Hand-letter “TOMBSTONE” at the top. No other text.

2. Create a reference comic sheet

PAGE 1 — Title bar: “THE CONVERGENCE ROOM • 1 • THE WITNESS”. 2×2 grid.

PANEL 1 (wide, flashback, wavy border): San Francisco in 1906 after the earthquake: collapsed brick buildings, smoke, a tilted streetlamp. A small black cat scurries out of a broken doorway.
CAPTION: “SAN FRANCISCO, 1906. THE EARTHQUAKE DIDN’T ASK ANYONE’S PERMISSION.”

PANEL 2 (medium): The same cat, now TOMBSTONE, a black marble statue sitting on the curved wooden bar of the Black Cat Bar, eyes closed, framed pictures and small brass lamps behind.
CAPTION: “LATER THEY CARVED ME IN BLACK MARBLE AND PUT ME ON A BAR. THE NAME’S TOMBSTONE.”

PANEL 3 (wide, exterior, night): The red-brick clubhouse on Nob Hill from the reference photo, arched windows glowing; a cable car passes; anonymous guests in evening wear climb the steps.
CAPTION: “TRADITION SAYS ANYTHING SAID BEHIND THE BLACK CAT IS OFF THE RECORD.”

PANEL 4 (close-up): TOMBSTONE’s face fills the panel, eyes narrowed, tiny silhouettes of guests reflected in his glossy marble.
CAPTION: “ON 27 OCTOBER 2026, TWENTY LEADERS KEPT SAYING THE SAME THINGS. I’D NEVER HEARD THAT BEFORE.”

Strict black ink on white. Text exactly as written; no other text.

3. Create comic styles for each person

Character model sheet for ANKOR, one of the four recognisable hosts in this comic.
Image 1 is the STYLE KEY: match its ink style, line weight and lettering exactly.
Image 2 is a photo of ANKOR: use it only for his likeness.

Draw ANKOR in the STYLE KEY’s black-and-white brush-and-ink style, on a white background, square:
(a) front head-and-shoulders, neutral; (b) three-quarter view, mid-sentence, one hand gesturing; (c) full figure standing, holding a wine glass; (d) close-up, laughing.
Keep the features that make him recognisable in the photo: slim face, short dark side-parted hair, clean-shaven, calm expression.
Costume for the whole comic: solid-black suit, white open-collar shirt, no tie.
Hand-letter “ANKOR” under the drawings. No other text.

Step 4: Generate and correct the images

This is a slow process. For each comic page, it gave me a generation prompt and a verification prompt. For example:

Here's the page 2 generation prompt

PAGE 2 — Title bar: “THE CONVERGENCE ROOM • 2 • STEAK, FISH OR VEGETARIAN”. 2×2 grid.

PANEL 1 (medium): ANKOR stands at the head of the dining room holding a wine glass, relaxed, guests at round tables listening. The walls behind him are bare: no logos.
ANKOR: “NO LOGOS ON THE WALLS TONIGHT. JUST DINNER, WINE, AND WHAT YOU’RE ACTUALLY SEEING.”
TAG: “— ANKOR, HOST”

PANEL 2 (medium close): RONNIE standing, hands open, mid-sentence, a slight knowing smile.
RONNIE: “PEOPLE DON’T RESIST CHANGE. THEY RESIST BEING CHANGED.”
TAG: “— RONNIE, MODERATOR”

PANEL 3 (close-up, top-down): A small printed menu card on the black tablecloth beside a wine glass. The card reads, exactly, as five short lines:
“TONIGHT’S MENU”
“STEAK: WHAT DOES ‘VALUE’ MEAN TO YOU?”
“FISH: EFFICIENCY, OR A BETTER CUSTOMER EXPERIENCE?”
“VEGETARIAN: WHAT’S STOPPING YOU FROM SCALING?”
“SIDES: ONE THING IN, ONE THING OUT.”
CAPTION: “PICK ONE QUESTION, LIKE A MAIN COURSE.”

PANEL 4 (medium, comic beat): A waiter approaches with a tray of covered plates; an anonymous guest at the table blocks him with a raised palm. TOMBSTONE watches from the bar in the corner, deadpan.
GUEST: “NOT UNTIL EVERYONE HAS SPOKEN!”
CAPTION: “ONE GUEST WENT TO WAR WITH THE KITCHEN. THE KITCHEN BLINKED.”

Strict black ink on white; hosts match the CAST SHEET. Text exactly as written.

Here's the verification prompt

Proofread the image you just made against my script for this page. List every balloon, caption, tag or sign whose text differs by even one character, any extra text or logo, any colour, any host who doesn’t match the CAST SHEET, and any guest who looks like a host. Don’t regenerate yet.

It turns out that the image DOES have a major issue. According to the review, the waiter in Panel 4 looks rather like Ankor, and fixing that is the highest priority correction.

Now, reading every verification report is painful. So I told Codex:

I’m following the instructions from prompts.md and have roughly completed up to step C2.

Continue and complete the rest. The tab is open on the browser.

Now this was genius, because Codex read ChatGPT’s output, send this follow-up prompt:

Edit only panel 4 of page 2: make the waiter a distinct anonymous older woman, not resembling Ankor or any host. Preserve the tray, scene, all lettering and every other panel unchanged.

… and then:

Ran the verification prompt again (2 minutes)
Read it, was satisfied
Moved to generating the cover page (3-5 minutes)
Ran the verification prompt (2 minutes)
Read it, and asked ChatGPT to remove the lettering and logos on bottle labels (3-5 minutes)
… and so on.

I asked it to run in parallel across four chats. So, while I was reading Asterix, an entire comic story was built.

Basically:

Claude wrote the prompts.
ChatGPT ran the prompts.
Codex got ChatGPT to run the prompts.
I read Asterix and worked on other things.

28 Ways I Review Agents

Fri, 06 Nov 2026 10:13:29 +0800

Using the same approach as specification and verification I looked at how I review agents, i.e. what I comment on after they share an output.

This page features the 28 review techniques I use. Some interesting ones:

Show your style: I tell it why it doesn’t sound like me, and how I would have written it. Real example: “Still not in my style - revise more in line with these principles.”
Pause and resume: Tell it not to exceed the usage limits. Real example: “We are at 96% usage of the 5h limit - so pause now and we can resume in a few hours.”
More like this: Tell it what you liked and go in that direction. Real example: “Looks good. Give me 10 more.”
Take the best of each: Give it other models’ output and pick the best from each. Real example: “Here is ChatGPT’s response. Take what’s better, drop what’s worse, follow up on new ideas this gives you, and revise your response.”
Say “I meant…”: Clarify your real intent. “Sorry, I meant round off the 95% activity window in EACH namespace.”

These are based on ~6.7K real responses. Skim through them and if you spot something interesting, test it out. If it works for you, share it with others.

Why are flights productive

Thu, 05 Nov 2026 18:02:41 +0530

This Thursday, I flew Indigo from Singapore to Chennai. I had no WiFi and got a lot of (slightly difficult) reading done, such as:

Machines of Loving Grace
Against the Personal Theory of Agents
Paul Graham’s The Top Idea in Your Mind
How I Built the Most AI-Pilled Company in the World
Catching up on Simon Willison’s blog
Actions from a long pending book cluster summary (The Beginning of Infinity, Conjectures and Refutations, The Logic of Scientific Discovery, The Fabric of Reality, and Enlightenment Now).
Several tweets, blog posts, and ChatGPT / Claude chats.
I also drafted prompts to run later.

In a disconnected flight, I’m less distracted because:

I don’t need to check on whether an agent completed.
I can’t open Hacker News if an essay is hard to read.
I can’t click on a link in the essay.

That forces me to absorb harder stuff.

Last Thursday, I flew Emirates from Brussels to Dubai. I had good (free) WiFi on the flight. Strangely, I still got a lot of things done:

Created an Asterix browser
Created a comic story using real faces
Ran an agent-swarm experiment
Improved video embedding on my blog
Published a few blog posts

It’s a different kind of productivity - doing rather than understanding - but still felt more productive. I think that’s because:

I didn’t have any meetings - unlike typical weekdays.
I can’t snack or watch TV - unlike typical holidays.

I tried this by sitting in libraries, at office on holidays, at cafes and client offices. It’s not the same. There’s something about the forced “You can’t get up for 7 hours” that makes me a bit more productive.

Your mileage may vary. But for me, flights are productive. I thought it was because there was no WiFi but it may be because I can’t get up.

A day later, I watched myself as I got distracted. Here’s what I noticed.

When tasks are running in parallel, I get distracted. “What still running? Did it complete?”
- Grouping all live tasks into a bucket - e.g. a tab group - helps. I’m peaceful there’s one place I can go check.
- Hiding the live task list helps - e.g. the browser vertical tab list and VS Code terminals. It’s more work to check what’s running.
When something updates on the screen, I get distracted. “What’s that?”
- Turning off things that update the screen helps, e.g. live sessions, notifications, even the running icons in browser tab / VS Code terminal titles
- Two screens have more distraction opportunities than one.

When do I get distracted?

When I get ideas (meetings, reading news), I get distracted. That’s a good thing - but I need to complete them and that requires delegation, prioritization, etc.
After finishing something (submitting a prompt, committing code), I get distracted. That’s a good thing - there’s no point waiting for an agent to complete, usually.
While doing other things (exercise, TV, brushing teeth, etc.) I get distracted. Up to a point, that’s OK.
When work is hard (complex, boring), I get distracted. That’s a bad thing - and precisely when I need to focus.

This is not structurally different from my fairly detailed experiments in 2022-23. I haven’t consciously applied those learnings to AI agents, I guess.

Worth practicing.

What Changed Your Mind?

Wed, 04 Nov 2026 14:00:00 +0530

I conducted a session on Wed, 04 Nov 2026 at ED5518: Data Visualization for Engineers - Department of Engineering Design, IIT Madras, Chennai.

Speakers: Anand S

Summary: A live-lab class on testing visualizations and phone apps: commit to an answer before discussing, check that a chart is true and that it works for its audience, use AI as a different kind of reviewer rather than a stand-in for people, and ask busy people “A or B?”

Here’s the link to the session

Links:

Transcript
Audio (81 min)

Innovations 2 Nov 2026

Tue, 03 Nov 2026 14:26:39 +0530

The Innovation team has been experimenting with 3D interaction, audio, video and AI agents. Three things struck me:

We can manipulate a 3D molecule with our hands, or try different furniture in a virtual room, using a browser.
Sound search can find things a transcript never records. And a small specialist model can diagnose a machine at a tiny fraction of an audio LLM’s cost.
AI can help us improve videos and business decisions when its suggestions are checked against evidence, not just written convincingly.

New ways to interact and create

Hold a 3D molecule in your hands (Sujal). At our Innovation Team meeting, I used hand gestures in front of my camera to manipulate a colourful 3D ATP synthase molecule: pluck, slice, scale, rebuild and explore its parts. The browser recognizes the gestures; a paired phone can also control the scene. It is related to Sujal’s remote 3D pottery experiment.

Uses: Science education, pharma/medtech training, virtual exhibits, remote product demonstrations.
Ideas: Could a biology teacher pull apart a molecule while explaining it? Could a field engineer demonstrate how to assemble a part without shipping the equipment? The gestures take practice, and this is a visual learning interface, not a validated physical simulation.

Rearrange a room, not just render it (Sujal). Roomshift lets you move furniture through a 3D room, swap pieces, compare layouts and walk inside. You can bring your own 3D model; the important difference from image generation is that the objects remain editable.

Uses: Real estate, furniture retail, interior design, offices and facilities.
Ideas: Could customers try a product in their own space before buying? Could we test hotel or hospital layouts with staff? Measurements are estimates until checked.

Find where a video might lose viewers (Sujal). Choose a demo video, click Analyze, and inspect a timeline showing cuts, speech, silence, visual formats and potential problems. In the Uber Eats Super Bowl ad, it suggests introducing the brand sooner and flags a static four-second stretch around 1:22. Every suggestion links to measured evidence, rather than being a free-form AI opinion.

Uses: Marketing, video production, learning-content teams and media publishers.
Ideas: What if every training video or campaign got an automatic editorial first pass? The tool flags possible attention problems; it does not measure actual audience retention or predict views.

Search and understand sound, not just speech

Find sounds that transcripts cannot search (Atharva). Search 7,221 clips for laughter, whispers, musical styles and conversations without using transcripts. In the live demo, CLAP found 8 of the top 10 human-labelled laughing clips. Atharva found CLAP stronger for recognizable sound events like coughing, sneezing and laughter, while Gemma is stronger for conversational scenes like friends joking about football. The demo can use both.

Uses: Broadcast and podcast archives, contact-centre QA, accessibility, video tagging.
Ideas: Search an archive for applause, sirens, background music or a change in tone. Could we enrich transcripts with a second index of everything audible that isn’t said? The hosted demo has preset searches; arbitrary new queries need its local backend.

Diagnose a compressor from one second of sound (Jivraj). On recordings of one compressor in eight states, a specialist industrial-audio model, FISHER, identified the correct fault 97.4% of the time after seeing five examples of each state. Simple signal processing reached 96%; Gemini 3.1 Pro with no examples scored 16.3%. Estimated analysis cost: $0.10–$0.27 per compressor per month for the cheaper approaches, versus $254 for the zero-shot LLM.

Uses: Manufacturing, predictive maintenance, industrial monitoring.
Ideas: Before deploying an expensive general model, try a microphone and a few labelled fault recordings. These results are from one machine in one lab; transfer to a second machine remains untested.

Can a phone hear your password? (Eswar). After training on the sounds of 25 presses per key, a small model identified 93% of unseen mixed keypresses from a nearby phone microphone. The page lets you listen to a word and see what the model inferred.

Uses: Security awareness, endpoint and microphone risk assessments.
Ideas: What can an innocent microphone reveal? It is device-specific: tested on different devices, accuracy fell close to chance.

Can we detect a cloned voice? (Eswar). A WavLM detector classified 94% of 604 held-out real/fake clips correctly, including simulated phone audio. But it caught only one of three newer voice clones when passed through a simulated phone line.

Uses: Banking fraud, contact centres and identity checks.
Ideas: A suspicious voice score could trigger callback verification, not automatic rejection. The newer-clone trial is tiny, and false alarms become costly when most callers are genuine.

The system around the model matters

Search 10,469 financial filings: simple versus agentic (Hritik). On 50 questions with the same Gemini 3.8 Flash model, agentic search achieved 81% answer correctness against 54.9% for BM25. But it cost about 3.2× more, took 178 seconds rather than 23, and had slightly lower faithfulness (87.6% versus 92.1%).

Uses: Asset managers, investment research, enterprise knowledge search.
Ideas: Don’t choose a RAG architecture from a single accuracy score. Compare correctness, citations, latency and cost on the client’s questions.

Agents that cannot give themselves permission (Eswar). In a simulated bank, eight agents investigate a trader’s $2 million discrepancy. Every request goes through a code-enforced gate checking both agent permissions and the human’s entitlement. An auditor probes the rules; governance proposes changes; Compliance approves them.

Uses: Banking, healthcare, agent security and compliance.
Ideas: Let AI discover and propose fixes, but keep authorization and enforcement outside the model. The replay is synthetic, not a production security certification.

One experiment for education

Can an AI be wrong in exactly the way a student is? (Atharva). Across almost 68,000 questions to nine models, prompting a model to hold a specific maths misconception can make it choose the corresponding wrong answer. Reasoning often helps it make the intended mistake.

Uses: Assessment publishers and adaptive-learning platforms.
Ideas: Let simulated students test multiple-choice options for misconceptions they do or don’t catch, before giving them to real students. Their behaviour still needs checking against actual student choices.

Also worth exploring: Reproducible LLM-answer replay (Eswar)—it saves and replays the first answer, rather than controlling model randomness—and news grouped by headlines, patterns and underlying forces (Atharva), which could be adapted into a client-specific industry briefing.

These are experiments. If one sparks an idea or connects to a client problem, please let me know.

52 Ways I Specify and Verify Agents

Mon, 02 Nov 2026 20:20:52 +0530

Since agents do most of the work, specification (telling it what to do) and verification (making sure it did it right) is more important than execution.

This is what I’m often asked about. Of the ~600 questions people asked me this year, half were deciding what to do / ask for and every sixth question was about checking if it worked. I checked.

For example:

“What should we ask agents to do when we only have a dataset and no clear use case?”

Or:

“How can I verify agent work when I do not know the domain well enough to judge it?”

But my advice usually has 3 problems.

It’s incomplete. I share 2-3 techniques, but I use dozens.
It’s unstructured. People already know some techniques. I add more to the pile, but there’s no clear pattern.
It’s unprioritized. Some techniques work better than others. I rarely explain which.

Luckily, ChatGPT, Claude, Codex, etc. allow us to export our conversations. So I searched over 11,000 prompts I’ve given agents this year to extract the top techniques I use to specify and verify work. I estimated how often I use them and ranked them by how useful I think they are. Here’s the result.

This page features the 30 prompting approaches I’ve used and the 22 verification approaches I’ve used, prioritized based on frequency, impact, and effectiveness.

For example, my top prompting techniques are:

Say where and how to save it: That way, I know where the result is and what to do with it next. Real example: “Create a Word doc which has images of 73 plants, each with a heading with the plant’s name (common name in English or Hindi). Save it on @LocalMCP as ~/Downloads/plant-images.docx.”
Give it the output format. That way, it knows what to create. Real example: “I have uploaded a market report, from last week. Create exactly the same report in exactly the same format, but for today.”
Point it to the right source. That way, it reads the material I meant, not something similar. Real example: “Using Local MCP, read prompts/gwslog.md and give me the feature list I would want from gwslog.py based on the features I’ve requested for recent scripts and how I’ve used them (see prompts/*.md, actual scripts, tmp.fish-history)”

You probably already do these. Here are some from the end of my list - and they’re more interesting. Like:

Describe the style. Real example: “Write like Malcolm Gladwell.” Or “Podcast style, no formatting.” Very helpful in making the output easier for you to read.
Share what you really did. Real example: “FYI - no response required. I sent this with almost no changes after reviewing. It’s a reasonable approach.” At the end of a chat, I tell the agent what I actually did or changed. It helps when I continue the chat later - and gives me useful data for future analysis.
Just say “Continue”. I ask the agent to plan, then execute the next step when I say “Continue”. This lets me review the plan before it acts, and break big tasks into smaller steps.

My top verification techniques are:

Run it for real. Tell it to run realistic test cases and show the results. Code that looks right may still fail. Real example: “You can’t write to it directly, but create a copy, run and test, verify that it’s working as intended in every way, revise as required, and let me download the file.”
Stage before it’s irreversible. Let it test or prepare a change before applying it for real. Real example: “Draft only. NEVER send or modify email.” Or, “OK, proceed with this, deploy to exam-dev and test it.”
Check the checker. Is your test case correct in the first place? Are you checking the right behaviors? Real example: “Before fixing, make sure the test is correct and the code is what needs fixing. then fix.”

Again, most people do at least #1 and #2. #3 was initially a surprise to me - but checking the verifier has quite a high ROI. Otherwise, the agent might actually break correct code.

But as before, the ones near the end are more interesting. Like:

Test with alternatives. Make sure the result doesn’t change when you change the inputs a little bit. Real example: “Try a few more prompt variations on a sample and see if any are promising and add the most promising variation to the result set.”
Show actual text. You can spot errors instantly with text you recognize or verify. Real example: “Show key snippets of actual prompts & results verbatim in code blocks.”
Flag the guesses. That way, I know what to check. Real example: “Propose a resolution with evidence, or mark #UNRESOLVED and state what evidence would settle it.”
Fail loudly. Real example: “Work through @LocalMCP. First check that it works; stop if unavailable.” That makes sure it doesn’t silently assume something or try alternatives.

These are based on actual usage (over 10,000 prompts) and real data, so I guess that counts for something.

A good way to use this might be: start at the end and skim through. If you spot something, test it out. If it works for you, share it with others.

Better yet, paste this in your agent (if it can read past conversations):

Based on my past conversations, what techniques in https://files.s-anand.net/pages/verification-techniques/ (if any) are relevant to me AND necessary, e.g. because you tend to miss it?

List them, prioritized. Tell me how I can reuse those - without my having to tell you each time.

Should a Student Even Read the Exam?

Mon, 02 Nov 2026 20:00:00 +0800

I conducted a session on Mon, 02 Nov 2026 at TDS Course Ideation — Open Design Meeting, Sep 2026 - Remote.

Speakers: Anand S

Summary: In the fourth open TDS design meeting, the teachers ask what is left to teach once an agent can read and solve the exam — while doubling a spending cap students are forecasting, finding only eight drawings that fool every AI guess, and admitting a marketplace rewarded trading over usefulness.

Here’s the link to the session

Links:

Things I Learned - 01 Nov 2026

Sun, 01 Nov 2026 00:00:00 +0000

This week, I learned:

Enterprise software migration seems a massive medium term opportunity - Ethan Mollick. On the other hand, liquid software where anyone can tweak their own softare is emerging.
“But I’m not working for my agent, it works for me. It’s a computer. It’s not a person I have to keep busy and keep managing (and checking it’s managing it’s own agents).” - Ben Tossell
OpenAI announced that ChatGPT Chat (except Instant) will start getting metered from December 2026. This is a big deal for me because:
- I prefer ChatGPT Chat / Work to Codex because it can access my past conversations.
- Chat (in my ChatGPT $20 Plus) isn’t metered visibly and feels unlimited.
- By giving ChatGPT a plugin to control my computer, I get most of the benefits of Codex.
- But this free use will last only a month. I need to see how well GPT 6.1 Terra fills that gap - i.e. is it good enough to beat GPT 6 Sol and cheap enough that I can run my chats without worrying about exhausting limits?
Clairic is an extraction model from Cloudflare. It’s a Decision model (or System One model) like Jev: fast and lightweight. But this focuses on extraction of text, numbers, dates, even forecasts. It provides the confidence as well as input position of the text for reference.
Anthropic recommends using the main agent rather than subagents. Single agents on smart models work well unless you need independence #ForNow.
Ruby chocolate is the fourth variety of chocolate - alongside dark, milk, and white. (Might be a marketing claim than verified.) I tasted it. It’s fruity, slightly sour, not bad. My order of preference among the melted chocolates I tasted in a spoon at Chocolate Nation, Antwerp:
1. Gold. Caramel toffee butter cream . Delightful!
2. White W2. White chocolate. Something I usually enjoy. I was surprised I liked Gold more.
3. 823. Milk cocoa caramel chocolate. Slightly darker.
4. Power 41. 41% cocoa Caramel with less sugar.
5. Ruby. Fruity sour taste.
6. Arriba. Strong cocoa flavor.
7. 811. Bitter.
8. The rest were too bitter
I had no idea there was so much innovation in chocolates. Most of the ones I like are a century old. Good to see some revival in the last decade. ChatGPT
- 1880s: Truffles. chocolate that melts immediately in the mouth.
- 1907: Hershey’s Kisses.
- 1908: Toblerone.
- 1912: Pralines. chocolate shell with creamy filling.
- 1920: Chocolate-coated ice cream (like Magnum).
- 1923: Filled chocolate bars.
- 1923: Nougat-caramel chocolate bars (like Five Star, Mars).
- 1923: White chocolate.
- 1926: Chocolate syrup (Hershey’s).
- 1928: Peanut butter cups : Combined sweet chocolate with salty, savory peanut butter for a striking flavor and texture contrast.
- 1932: Chocolate oranges.
- 1935: Aerated chocolate with tiny air bubbles.
- 1935: Chocolate-coated wafers (KitKat).
- 1936: Malted chocolate balls (Maltesers)
- 1949: Soft-melting chocolate centers (Lindor)
- 1951: Chocolate-hazelnut spread (eventually Nutella)
- 1969: Spherical filled chocolates (Lindor)
- 1971: Chocolate eggs (Cadbury Creme Egg)
- 2012: Blond chocolate: white chocolate with biscuit and toffee notes.
- 2017: Ruby chocolate: pink, fruity, sour.
- 2021: Dubai chocolate: Pistachio cream, cruncy.
Looks like GPT 6.1 Terra is mostly as good as Sol but cheaper. I haven’t tested it yet.
On one hand, modern CSS has several new features. On the other hand, I’ve stopped writing code and delegated it to agents, so this is of no use to me. Instead, what I should be thinking about is: what new patterns does this enable, and what should I be asking for in prompts?
Ethan Mollick: “Ask yourself: “how do I make what I am doing even more ambitious?” Something to practice actively, given the growing abilities of agents.
Anthropic is comfortable that no single person has an architectural view of Claude Code. I guess architecture is also being taken over by agents. Addy Osmani: “No member of the team has a full mental model of Claude Code in their head, nor do we think anyone needs to. Rather, when anyone wants to learn about a piece of the codebase, they ask Claude to generate an explainer about that exact slice. It only takes a few minutes and the explainer is always up to date.”
Good advice from Thornton Bell
- Humans write for humans. Agents for agents.
- Be more ambitious.
- Give it more eyes - logs, observability, …
- Code on remote machines.
- Ask for 5 variations. (I’m not doing enough of this yet.)
- Use schedules and triggers. (I’m not doing this at all, so far.)
Codex introduced an experimental #ForNow --enable instant_interrupt flag that looks promising. It’s still steering, i.e. not instantly stopping any existing requests or tasks (hence wasting effort), but doesn’t wait for the entire response either. Not sure how it works but I just enabled it and will evaluate it. Code-mode yielding PR, model-response preemption PR.
In Codex you can @-mention another Codex session. That’ll give it access to an earlier session’s context. In the past, I copy-pasted session transcripts, saved them in files, or pointed Codex at the session directory. This is more convenient.
PixelBoard has pretty impressive cvoice typing. Google has a Gemini-powered “Rambler” voice recognition system (based on Gemini 3.5 Translate, with offline fallback) on Android phones, but not available via GBoard. PixelBoard enables that. ChatGPT

Mistakes I made

Week ending 01 Nov 2026

I said “LLMs can probably remember 20, 30 things. If you give them 1,000 things, they will almost certainly forget” while discussing spreadsheet classification.
Correction: LLMs do not have a fixed memory ceiling of 20–30 items. Modern long-context models can process thousands of examples or records within a single prompt. However, accuracy depends on the model, context length, task complexity, and number of items requiring attention. For classifying 1,000 records, I should compare direct prompting, batched classification, and embedding-based classification against a human-labeled sample rather than assume direct prompting will fail. Google: Long context · OpenAI: Optimizing LLM accuracy
MEDIUM · FALSE
I described drawing by tilting a mobile phone as “Your mobile phone’s tilting as a Gyroscope basically” while discussing an educational exercise.
Correction: A gyroscope measures angular velocity, not absolute tilt. Phone orientation is typically obtained through sensor fusion using the accelerometer, gyroscope, and sometimes magnetometer. An accelerometer can estimate tilt relative to gravity when the device is stationary; a gyroscope helps track rotation. For a browser-based tilt-drawing exercise, DeviceOrientationEvent exposes orientation angles, rather than raw gyroscope readings. Android: Motion sensors · MDN: Device orientation
LOW · OVERSTATED
I said “the entire suite of Autodesk, Microsoft, Corel, Adobe … every piece of software has been pretty much cloned within the span of 1 month” while discussing the declining value of traditional software.
Correction: AI-assisted open-source projects have rapidly reproduced substantial parts of several established applications, but the entire commercial suites have not been faithfully cloned in a month. For example, PhotoCraft explicitly describes itself as early-alpha software, not yet a professional Photoshop replacement, with missing tools, generative features, typography workflows, and plug-in compatibility. Repository existence, feature checklists, and working demonstrations do not establish feature parity, production reliability, or replacement readiness. I should describe this as rapid progress in recreating software capabilities, not wholesale replacement of entire product suites. PhotoCraft: Project status
HIGH · FALSE
I said “Figma themselves have ported the bulk of Figma … Figma itself is now available … inside ChatGPT” rather than merely being controlled through ChatGPT.
Correction: Figma has an official ChatGPT integration that can generate embedded previews and create editable FigJam diagrams and Figma Slides presentations. Figma’s documentation explicitly says the integration does not currently create files for its other products, including the main Figma Design editor. Users generally open generated files in Figma for further editing. This is a useful integration, not the bulk of Figma running inside ChatGPT. Figma: Use ChatGPT with Figma · OpenAI: Figma plugin
MEDIUM · FALSE
I said “Cost and quality, it trumps RAG like crazy” while recommending agentic search over retrieval-augmented generation.
Correction: Agentic search performed well in particular experiments, but it has not been demonstrated to universally outperform RAG on both cost and accuracy. Our earlier retrieval comparison involved a limited test set and unequal access to source material, while the larger agentic-search experiment was separate. These cannot establish general superiority. RAG costs also vary substantially with embedding models, index architecture, storage, query volumes, and infrastructure. I should present agentic search as a promising alternative and compare approaches using the same corpus, questions, accuracy criteria, and end-to-end costs. Retrieval benchmark · AWS: RAG cost considerations · OpenAI: Optimizing LLM accuracy
MEDIUM · OVERSTATED
I said “both Apple and Google for iOS and Android released a developer kit right up front” while comparing mobile app ecosystems with AI platforms.
Correction: Apple did not release a native iPhone SDK when the iPhone launched in June 2007. Initially, Apple supported third-party applications through web technologies. The native iPhone SDK was released in beta on 6 March 2008, more than eight months after the original iPhone launched. Android adopted a more developer-oriented strategy earlier in its rollout. The two platforms therefore followed different paths toward third-party applications. Apple: June 2007 web applications announcement · Apple: March 2008 iPhone SDK announcement
LOW · FALSE
I said “tricking AI is much harder” and “AI is not going to be fooled by bad designs of many sorts, but humans are” while proposing visualization assignments.
Correction: Multimodal models can also be fooled by misleading charts, and may perform substantially worse than humans on visual reasoning. In the NeurIPS 2025 ChartMuseum benchmark, humans achieved 93% accuracy versus 63% for the best evaluated model. Research on deceptive charts also finds that leading models remain vulnerable to manipulated axes, layouts, and visual encodings. I should design the exercise to discover which representations fool humans, AI, both, or neither, rather than assume AI is harder to fool. ChartMuseum, NeurIPS 2025 · VisDeception benchmark
MEDIUM · OVERSTATED
I said “the model can then evaluate the novelty factor through the research” while proposing that students invent a chart format nobody has created before.
Correction: An AI research agent can find precedents and compare a proposed chart with known visualization techniques, but its inability to find a similar chart does not establish that the design is unprecedented. Even a formal prior-art search is not conclusive proof of novelty. I should assess documented differences from the closest examples found, the usefulness of the design, and how thoroughly precedents were searched, rather than certify absolute originality. European Patent Office: Novelty and prior art
LOW · OVERSTATED
I said “AI can … check scientific claims by actually reproducing them” while circulating an experiment in scientific-paper reproduction.
Correction: The demonstrated agents test computational reproducibility: whether they can reconstruct reported numerical results from the original data and described methods. That is valuable evidence about analytical transparency and reproducibility, but it does not establish whether the underlying scientific claims are true. Reproducing results does not independently validate the data, experimental design, causal conclusions, or generalizability. The National Academies explicitly distinguishes computational reproducibility, replication with new data, and scientific validity. I should describe the demonstration as an automated computational-reproducibility check, not general scientific claim verification. National Academies: Reproducibility and Replicability in Science · Original experiment
MEDIUM · OVERSTATED

Questions I was asked

Week ending 01 Nov 2026

Question: Why are mobile phones particularly interesting for AI-built applications? Asked after I described how much more ambitious my experiments had become by connecting agents to Android phones.
Answer: Accelerometers and gyroscopes are remarkably good. Phones also have touchscreens, cameras that are much easier to point anywhere than laptop cameras, microphones, and a huge installed base. With a cable and USB debugging, agents can now access all of this remarkably easily.
Question: What will the commercial model be for software companies when their applications become plugins inside AI platforms? Asked while discussing Figma inside ChatGPT and how AI threatens conventional software businesses.
Answer: My guess is something like the iOS App Store. Software companies get distribution to a much larger audience and can eventually charge subscriptions or extra for AI-enabled features. But people can also build the software themselves with AI, so the plugin ecosystem is questionable. Enterprises may still pay for someone they can trust.
Question: Isn’t Responsible AI becoming a clichéd theme? Can we still show innovative demos under it? Asked after considering Responsible AI as the unifying story for a managerial CIO audience.
Answer: Yes, we can show demos under Responsible AI. At a managerial level, people will probably prefer safety as a topic over possibility. A developer audience would be different.
Question: How can we grade a data visualization project without spending hours manually evaluating subjective answers? Asked after a 40-student chart-analysis quiz had taken weeks to grade and while replacing part of the final examination with a project.
Answer: Give students small, testable challenges. For example, create a chart that humans and AI both interpret correctly, then one that fools humans but not AI. Use fixed questions, five-second human interpretation tests, and AI evaluation to measure the results. Students should use agents to do the work; our grading effort should be negligible.
Question: Should we let students choose their own applications when inventing new visualization formats? Asked while designing a project that tests creativity rather than routine chart generation.
Answer: Yes. Ask them to invent a chart format nobody has seen and explain the specific situation where it’s useful. Use AI deep research to judge whether it’s genuinely novel, then test its usefulness: can someone understand it in 30 seconds? Novelty alone isn’t enough.
Question: How should we evaluate whether a lightweight open-source video generation framework can replace an existing one? Asked after reproducing one HyperFrames benchmark task and comparing its output with Remotion and a golden reference.
Answer: Try three to five more, quite different tasks. Compare quality with blind judgments, how reliably an agent produces working videos without fixes, and cost and speed including setup overhead. Don’t rely entirely on HeyGen’s judge when the benchmark comes from HeyGen. Show where each approach wins and fails, then recommend when to use which. No need to run all 50 tasks.

Drawing humans only CAPTCHAs

Sat, 31 Oct 2026 14:38:26 +0530

My TDS course asked students to draw objects in one minute. The drawing is sent to their classmates as well as three vision models. They score if:

Classmates guess it right
AI guesses it wrong

In other words, can you, in 1 minute, draw something that fools AI but not humans?

Here are some of the best results, when I ran it against GPT 6 Luna, GLM 5.3 Flash, and Deepseek v4.1 Flash, which are really cheap and good vision models.

How to fool AI

AI was convinced that this ladder was a saw!

Most people guessed this as ladybird (lady + bird) but Luna said “butterfly”, Deepseek said “dragonfly”, and GLM said “butterfly” but did, once, guess ladybird.

Everyone guessed this as donkey. AI guessed sheep, cow, horse, pig, and even turkey.

Here’s a good CAPTCHA idea. All the humans guessed this as lamp. AI thought hammock, occasionally table or chair.

Everyone guessed this single stroke as a caterpillar. But most AI thought it was a snail.

Everyone guessed this single stroke as medal AI thought it was a whistle, tennis racket or shuttlecock.

But the one that puzzles me the most is: why did AI struggle so much with this mango? Humans got it right. AI guessed pear - even banana.

How to fool humans

This is the reverse category. AI guessed it right away but people couldn’t.

No, this does NOT look like an owl to me. People guessed: pigeon, penguin, flamingo, crow, eagle - but AI guessed owl every time.

This does NOT look like a seahorse to me and people guessed: jellyfish, whale, fish, lobster, whale. But AI guessed seahorse every time.

Horses are really tough to draw. People guessed: buffalo, buffalo, donkey, goat, cow. AI guessed horse almost perfectly.

How to fool everyone

And then there are images that EVERYONE gets wrong.

That’s supposed to be a turkey.

This is a zebra. Not a camel, giraffe, elephant, or even a kangaroo.

This is a guitar. Not a middle finger.

Does a lemon have a stem and leaf, usually?

Good attempt at shark that NO one guessed. Everyone said fish.

It’s REALLY hard to differentiate cow from buffalo.

And finally, I drew a jellyfish like a helicopter when I was testing because I didn’t see the word jellyfish!

Humans are slightly better than AI at guessing humans’ drawings. The advantage is ~3.5%. Strikingly, 8 drawings were understood by every human but missed by every AI. Only 2 drawings were the reverse.

GLM 5.3 guesses very similar to humans (75%), while DeepSeek 4.1 Flash is next closest (73%), and GPT 6 Luna the furthest of the three (68%).

↓ Humans / AI →	Right	Partly right	Wrong
Right	368	112	8
Partly right	104	196	30
Wrong	2	15	31

No AI Homework for Kids

Sat, 31 Oct 2026 09:46:00 +0530

I conducted a workshop last weekend for working professionals and had assigned homework as part of that. We reviewed their homework at the start of the class, and I noticed that the majority of them had clear AI smells.

I let the participants know that using AI is perfectly fine, especially in an AI class, and asked how many of them used AI to write the final output, All, except one, said “Yes” - a bit of a surprise to me since it included some output that I thought was manually written.

Then I asked them how many would allow their children to do the same, that is, to submit AI output as their homework. A clear majority of them said no.

	YES, kids should use AI	No, kids shouldn’t
YES, I used AI	7	16
NO, I didn’t	0	1

(3 others who didn’t submit their homework also said “No” to kids using AI for homework.)

I mentioned this to my daughter, who nodded wisely and said, “Well, at least they’re self-aware!”

Code on remote machines

Thu, 29 Oct 2026 17:23:48 +0100

Thornston Bell’s How I work with Agents suggests actively using remote machines for coding. I’d struggled with this for many months, and finally managed to set up my Windows laptop for remote coding.

How?

Tell my wife to tell the maid to never switch it off. (It doesn’t have battery.)
Tell ChatGPT to install Remote Desktop Commander - and schedule it to run on boot. Now agents can use it.
Tell ChatGPT to install Tailscale so I can SSH into it and copy files.
Tell ChatGPT to create a sync-plutonium.sh that pushes files from my Linux system (graphene) to the Windows machine (Plutonium).

That last step was the critical enabler. This means that most of the files on my regular laptop is also on the home laptop, and it has the same context.

This is proving to be quite a big unlock!

For example, I landed in Brussels on Monday. The immigration queue for took about 2 hours and 15 minutes for me - as a non-EU citizen. I used this time to:

Install Taildrive on my home laptop and Material Files on my phone, so I can edit files in Singapore from Brussels.
Draft an email and blog post about our innovation team’s work last week. Since all the transcripts and repos were synced, ChatGPT could find up everything, create screenshots, and explain it the way I wanted it.
Run an agent swarm experiment - to see if 4 Codex agents running GPT 6 Luna would, if allowed to do what they want, do something interesting. They didn’t - they decided not to waste tokens. When pushed, they fumbled and formed two groups and built two tools.
Ideate how to structure a market economy that grows in value when students do useful work - as part of my TDS course. Since it had the code base for my entire course and evaluation, ChatGPT could give me relevant ideas.
Read Anthropic’s code review process.

On the way back, the emigration queue was only 1 hour and 30 minutes. I used this time to:

Make my Calvin and Hobbes browser mobile-friendly. My blog was already cloned on my home laptop.
Create an annotated Asterix comic viewer. The comics were available on the home laptop. So was Python (for OCR).
Mine all my data and documents to understand my preferences - since almost all my docs are on the home laptop.
Complete email requests and tasks from several colleagues - since I’d give it access to my email via gws.

Increasingly, I find myself comfortable not taking my laptop when I go out. I can talk to my phone and get stuff done - and that might be more efficient than me typing it out on my laptop.

Reading Asterix from Brussels to Dubai via Singapore

Thu, 29 Oct 2026 15:35:52 +0100

I think it’s a miracle to see the sheer number of technologies that can come together with AI agents orchestrating them.

I’m right now on a flight from Brussels to Dubai, and the Wi-Fi works. So, I’m dictating this blog post via PixelBoard, an AI-coded port of Gboard that lets me use Gemini for voice typing on Android (which is quite good!).

I’m also browsing my Asterix comic book collection. These are on my Windows laptop in Singapore, which I’ve connected via Tailscale. I have no idea how to configure Tailscale, and ChatGPT did that for me. BTW, Tailscale is one of those services that I didn’t see a use for until I used it, and now will never turn it off. It creates a VPN between my phone, traveling laptop and home laptop, so I can browse files, SSH into each other, run HTTP servers, etc. All securely.

I told ChatGPT to create a comic book reader I could use on my mobile; to use OCR to detect the panels bounding boxes; to annotate the jokes (like Crismusbonus is Christmas Bonus) and facts (like Vercingetorix’s Arverniactually defeated Caeser at Gergovia in 52 BC, etc.); and run it on a server.

Video

It’s phenomenal that, sitting on a flight, I can connect through my voice to my laptop in Singapore and run an OCR engine and have it serve comics purely for my amusement mid-air. All with technology that I couldn’t have manipulated by myself. And all I need is the imagination to ask for it.

Innovations 26 Oct 2026

Tue, 27 Oct 2026 07:02:14 +0100

The Innovation team has built a few more demos since the last update. Three interesting learnings:

We can speak in one language and be heard in another, live, in a normal Google Meet call.
A person in another city can use their hands and a phone to sculpt a 3D object on your laptop.
AI can search what was heard, not just said, and can check scientific claims by actually reproducing them.

New capabilities

Translate a live meeting into almost any language (Hritik). At Friday’s VIA talk, I spoke and the audience heard Filipino, Hindi, Tamil and Telugu with a short delay. Hritik built a desktop app that creates a virtual microphone and speaker for Meet. It translates both ways: what others hear from you, and what you hear from them. It took him a day or two to vibe code, and comes as a Windows, Mac or Linux app. I estimated about 0.6 US cents a minute in my trial.

Uses: Global customer service, multilingual client discovery, remote consulting and workshops.
Ideas: One attendee immediately thought of discovery with a Spanish-speaking client. Could service teams work across language boundaries? Could every listener choose their own language? There is some voice delay, and virtual audio devices still need setup on each computer.

Sculpt a 3D pot with your hands, from anywhere (Sujal). Sujal used his laptop camera to track his hands and shape a pot, then Jivraj used his phone from Hyderabad to manipulate the same 3D object remotely. The result can be exported as a 3D model or JSON. We saw it work live, with some lag.

Uses: Product design, education, remote training, rehabilitation and industrial operations.
Ideas: If gestures can control a model remotely, can an expert teach somebody to sculpt, assemble machinery or lift equipment safely without being there? Can the same phone sensors guide a real machine? Sujal also built 3D maths and geometry controls. The example below was sculpted in the web page with its mouse fallback; the live hand-tracking demo was at VIA.

Search audio for things nobody said (Atharva). An embedding model searches over 7,000 clips for sounds like whispering, laughter, an angry tone or a jazz instrument, without a transcript. Click “Whispering”: all ten displayed matches have whisper labels. Atharva compared CLAP with EmbeddingGemma 2; neither model is best for every sound.

Uses: Call centres, meeting recordings, media archives and music libraries.
Ideas: Find meetings where people laughed or argued, or customer calls where the tone changed, even if the transcript looks innocuous. The hosted page has working sample searches; free-text searches and playback require its local server.

Checking what the results say

Can AI independently reproduce a scientific paper? (Jivraj). The agents get a paper’s Methods and publicly deposited data, but not the original code or results. They write code to rebuild a table, and a fixed script checks every number. On six papers from Elsevier, Springer Nature and Wiley, Sonnet 5.5 reproduced 77% of 495 numbers within 2%, versus 71% for Opus 5.5, at about $1.11 versus $2.56.

Uses: Scientific publishers, research-integrity teams and editors.
Ideas: What if a paper came with a reproducibility report that pinpoints which methodological choices were missing? Six selected papers and mostly one attempt each are only a pilot, but the individual failures are interesting.

How often does Veeva change a release-risk rating? (Eswar). Across 11 Veeva Vault releases with comparable old and new ratings, 91 of 2,547 features (3.6%) were re-rated. Just 1.9% changed whether a feature needed testing.

Uses: Pharma software validation, regulated technology and QA.
Ideas: Track what vendors initially say and what they later correct. Could validation teams automatically catch changes that actually affect testing?

Let agents handle the routine cases

Show a shipping manager only what needs a decision (Eswar). The system combines Rhine water levels with dated carrier notices, converts notices to versioned rules, and checks shipments. On the selected day, 26 shipments need a human, while 27 are handled without a screen. The water levels and carrier notices are real; the 420 shipments are invented.

Uses: Logistics, shipping, supply chains and exception management.
Ideas: Perhaps a good operations dashboard should hide everything that is going right. Surface exceptions with the evidence, the rule version and the decision history. Try “Replay the season”.

And more

A pharma tablet-plant twin (Eswar). The factory demo from last week now follows a tablet batch through production, showing where an agent can act and where QC approval is mandatory. A synthetic, guided demo.
Send messages between devices using sound (Jivraj). Laptop and phone communicating through speakers and microphones, without a network. Interesting for offline environments; still an experiment.
An AI meeting stand-in (Eswar). An explicitly labelled AI can answer some questions from your notes and show slides. Prototype, not an unattended meeting assistant.
Eight AI delegates negotiate a budget (Eswar). Their private instructions and secret votes make a fun workshop discussion about what to delegate to agents.

If you get any ideas from these or want us to try something for a client, please let me know.

A Mobile Voice Choir

Mon, 26 Oct 2026 05:06:29 +0100

I’ve been playing around with mobile sensors recently and one such experiment is the voice choir.

Here’s how it works.

I’m the presenter. I visit voice choir, create a room, and show the QR code.
People see the QR code (via screen-sharing or projector) and open on their mobile phones to join my room.
I press “Start”. They hear a tone and they should hum at the same pitch.
They can see if their pitch is above or below and adjust. I get a data stream of their pitch.
Over time, we see if we can hum at the same pitch, or how close we converge.
I record this as a visualization.

We played this in Straive across about 20 people who joined the room. Here’s the video:

Video

What surprised me most was how little convergence there was even after 45 seconds. People were still humming all over the place.

In the early days of Gramener, I learned that there are some people who find it hard to pick up a “design” sense. Even after repeated guidance, I found them violating rules like keeping edges aligned. Research also suggests that some people find it hard to pick up mathematics - and that the population may be bi-modal (i.e. some people get it, some just don’t.)

Maybe this true for music as well?

I think it’s cool that we can test this: get 100 students into this room, have them hum a tone, and see if the alignment / convergence is bimodal!