<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:media="http://search.yahoo.com/mrss/"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>Towards AI</title>
	<atom:link href="https://towardsai.com/feed" rel="self" type="application/rss+xml" />
	<link>https://towardsai.com</link>
	<description>Making AI accessible to all</description>
	<lastBuildDate>Tue, 06 Oct 2026 08:33:18 +0000</lastBuildDate>
	<language>en-US</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=6.9.10</generator>

<image>
	<url>https://towardsai.com/wp-content/uploads/2019/05/cropped-towards-ai-square-circle-png-32x32.png</url>
	<title>Towards AI</title>
	<link>https://towardsai.com</link>
	<width>32</width>
	<height>32</height>
</image> 
	<item>
		<title>The 3 Files That Make Claude Code Much Smarter</title>
		<link>https://towardsai.com/p/machine-learning/the-3-files-that-make-claude-code-much-smarter</link>
		
		<dc:creator><![CDATA[Codebook Fusion]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 08:06:52 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/the-3-files-that-make-claude-code-much-smarter</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Codebook Fusion Originally published on Towards AI. Learn the three files that make Claude Code smarter: CLAUDE.md, settings.json and SKILL.md. Real examples, setup steps, and honest tradeoffs. For my first few weeks with Claude Code, I did something embarrassing. I typed the same instructions at the start of every session. Use pnpm, not npm. Run the tests before you say you’re done. Leave the migrations folder alone. Then I got annoyed when Claude drifted off the rules. IMAGE GENERATED BY CHATGPTThe article explains that Claude Code improves dramatically when you stop re-prompting rules and instead store them in three purpose-built files: CLAUDE.md (project context and guidance), settings.json (permissions and enforced hooks so Claude can’t “talk its way around” constraints), and SKILL.md (repeatable workflows you can run by name, optionally bundling supporting files). It walks through what to put in each file, key tradeoffs (e.g., guidance vs enforced rules, keeping files small, starting permissions narrow), a quick 15-minute setup checklist, common failure modes like bloating CLAUDE.md, and an FAQ clarifying where these files live and how skills work. The takeaway is to write all three files small and focused so your workflow stays consistent with far less manual repetition. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*fK-ZnzOGsC6AYqvc0MwxDg.png" medium="image"></media:content>
            	</item>
		<item>
		<title>System 1 (Jev) Models: Faster and Cheaper Proxy to Frontier Models — With a Working Router</title>
		<link>https://towardsai.com/p/machine-learning/system-1-jev-models-faster-and-cheaper-proxy-to-frontier-models-with-a-working-router</link>
		
		<dc:creator><![CDATA[Muhammad Soliman]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:33:09 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/system-1-jev-models-faster-and-cheaper-proxy-to-frontier-models-with-a-working-router</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Muhammad Soliman Originally published on Towards AI. Jev, Laya, Kev and Nimble in one place — plus a short, open proof of concept that puts one in front of a LiteLLM proxy and picks the model for every request. Ask your coding agent to rename a variable, and it will load a few hundred billion parameters to do it. Is it worth to pass everything to LLM as is and waste hundreds of thousands of tokens for simple questions/routing/decisions — The Most Expensive Question You Ask an AI Is “Yes or No” and System 1 model was built to save alot of respond faster. The article explains why “System 1” (decision models) can replace expensive full LLM calls for routing and other typed decisions, contrasting it with System 2 token-generating models. It introduces how these models work (typed questions answered in one forward pass, returning probabilities with no generated tokens), covers the new availability of four options (Jev, Laya, Kev, and Nimble), and discusses how to read benchmarks carefully—these models are meant as bases to specialize via small fine-tuning sets, keep option sets small, calibrate confidence, and match hardware to the latency/claims. It then shows practical integration: using a classifier to choose among backend models (e.g., haiku/sonnet/opus) before the request leaves the machine, demonstrating a proof-of-concept tier router implemented as a LiteLLM callback that rewrites only an opt-in alias and fails open while logging decisions. Finally, it outlines where to take it next (measuring on real traffic, fine-tuning the classifier, adding guardrails in the same pass, quota awareness, and other proxy/gateway hooks) and concludes with the key takeaway about the strong economics and relatively small integration effort. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*35eS-NYncUVfD9Vi9-kdSQ.png" medium="image"></media:content>
            	</item>
		<item>
		<title>The Number That Matters in Cloudflare’s Clef System-One Model Isn’t 38.8 ms — Jev and Laya compared</title>
		<link>https://towardsai.com/p/machine-learning/the-number-that-matters-in-cloudflares-clef-system-one-model-isnt-38-8-ms-jev-and-laya-compared</link>
		
		<dc:creator><![CDATA[Muhammad Soliman]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:32:51 +0000</pubDate>
				<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/the-number-that-matters-in-cloudflares-clef-system-one-model-isnt-38-8-ms-jev-and-laya-compared</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Muhammad Soliman Originally published on Towards AI. The Number That Matters in Cloudflare’s Clef System-One Model Isn’t 38.8 ms — Jev and Laya compared Cloudflare just open-sourced Clef and Clef-flash, two “System One” decision models built on Qwen. The headline is speed. I went in to prove the fast one couldn’t say “none of the above.” I ran it on my laptop against Laya, and that hypothesis didn’t survive. What broke instead was more interesting. After explaining what decision models are (and how Clef fits into the “System One” routing-style protocol compatible with Jev), the article breaks down what’s actually new in Clef—especially its frozen Qwen backbone paired with a joint schema head for parallel scoring, along with explicit calibration objectives and multimodal inputs. The author then contrasts vendor benchmark claims with their own testing: Clef-flash’s routing performance is strong, but surprising weaknesses appear on out-of-scope intent detection (and execution latency varies heavily with hardware and option counts). Through three experiments—including CLINC150 OOS tests and real runs inside a browser agent—the author shows that safety depends on providing an explicit escape option and gating decisions using margin/escape probabilities rather than raw confidence alone. They also surface practical integration lessons (calibration differences across models, option-set size effects, drift from rendering differences, and why state/token cost matters more than decision-only latency), conclude with a recommended “gate you can ship” routing ladder, and caution readers to validate on their own agents because a crucial failure mode may not show up in public leaderboards. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*NCFdTab3SeySNaj00wD05Q.png" medium="image"></media:content>
            	</item>
		<item>
		<title>What Constrained Decoding Does That Prompting Never Can</title>
		<link>https://towardsai.com/p/machine-learning/what-constrained-decoding-does-that-prompting-never-can</link>
		
		<dc:creator><![CDATA[“The AI Engineer”]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:32:28 +0000</pubDate>
				<category><![CDATA[Artificial Intelligence]]></category>
		<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/what-constrained-decoding-does-that-prompting-never-can</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): “The AI Engineer” Originally published on Towards AI. Subtitle You asked the model for JSON. You wrote “return valid JSON only” in capital letters. You added an example. You added a second example. For 99 calls out of 100, it worked. image generated by GEMINIAfter the lead, the article explains the difference between prompting and constrained decoding: prompts shift probabilities via context but cannot force invalid tokens to zero, while constrained decoding applies a token mask to eliminate choices that would break a formal rule (e.g., a grammar or JSON schema). It walks through a simple masking mechanism and discusses real-world implementation challenges—especially that building the allowed-token mask is the hard, sometimes expensive part—highlighting research on faster preprocessing and GPU-friendly mask generation. The piece warns that “valid” doesn’t mean “likely” or “true,” since greedy local masking can distort the model’s distribution and may reduce output quality or reasoning performance under strict constraints; it mentions proposed fixes like ASAP and evidence from studies comparing constraint methods. Finally, it recommends practical patterns (e.g., think freely then constrain extraction, adjust strictness to the task), cautions against misleading metrics that become trivially green, and closes with guidance on when constrained decoding is worth using and how to interpret its guarantees. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*pKAH2aJ7ve4RGSeSxxP3_w.png" medium="image"></media:content>
            	</item>
		<item>
		<title>Why Is Microsoft Foundry’s Content Filter Blocking Legitimate Medical Questions?</title>
		<link>https://towardsai.com/p/machine-learning/why-is-microsoft-foundrys-content-filter-blocking-legitimate-medical-questions</link>
		
		<dc:creator><![CDATA[Dave R - Microsoft Azure &#38; AI MVP☁️]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:30:53 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/why-is-microsoft-foundrys-content-filter-blocking-legitimate-medical-questions</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Dave R &#124; Microsoft Azure &#038; AI MVP ☁️ Originally published on Towards AI. An AB-100 case study: read the guardrail annotation, find the category and severity, and relax one threshold while Self-harm stays strict. An oncologist asks a clinical notes assistant built on Microsoft Foundry to summarize the risks of a high-dose chemotherapy plan, and the app gets back an HTTP 400 with the error code content_filter. The same assistant still declines off-topic questions exactly as intended. This scenario is the false positive that teams running AI in medicine, law, or security eventually meet. Here&#39;s how to find the exact category and severity behind the block, and how to fix it with a custom guardrail while Self-harm stays at the strictest setting. Why Is Microsoft Foundry’s Content Filter Blocking Legitimate Medical QuestionsThe article walks through a four-step approach to fix a medical-system false positive in Microsoft Foundry: first, diagnose by reading guardrail annotations returned with API calls (including distinguishing input-block vs completion-block vs model refusal), then identify the specific harm category, severity, and whether the block occurred on the prompt or the completion; next, create a custom guardrail (since the default one is read-only) by adjusting only the implicated threshold on only the side that was blocked—while keeping Self-harm at the strictest setting for clinical safety and governance—along with maintaining other controls; finally, prove the fix by rerunning both your clinical test set and an approved self-harm test set, checking Azure Monitor metrics for the expected “harmful detected stays, blocked volume drops” signature, and auditing changes via the Activity log. It also explains why common distractors fail (blocklists only add blocks, the Asynchronous Filter changes timing rather than behavior, and Activity log is for control-plane auditing not content diagnosis), and emphasizes operational cautions like not using relaxed Self-harm thresholds for real clinical contexts without compliance review and avoiding privacy leaks when logging annotations. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*pHbJxQGBVdJRTg8wl2rNpg.png" medium="image"></media:content>
            	</item>
		<item>
		<title>Production RBAC, Cost Optimization, and Deployment Patterns for Cortex Agents</title>
		<link>https://towardsai.com/p/machine-learning/production-rbac-cost-optimization-and-deployment-patterns-for-cortex-agents</link>
		
		<dc:creator><![CDATA[Satish Kumar]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:30:42 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/production-rbac-cost-optimization-and-deployment-patterns-for-cortex-agents</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Satish Kumar Originally published on Towards AI. The developer-to-production pipeline for Snowflake’s Cortex Agent GA enhancements — Personal Database sandboxes, temporary agents, COPY GRANTS, and the cost math on Cortex Search suspension. Part 2 of 2 — Part 1: Your Cortex Agent Specification Is Visible to Every Role That Can Invoke It From Personal Database to Production: Building Secure, Governed &#38; Cost-Efficient Cortex Agents Part 1 covered why secure agents exist and how spec redaction actually works — including the ownership gotcha that silently breaks redaction when ACCOUNTADMIN owns the agent. This article covers the production deployment pipeline: how to develop, validate, deploy, and operate a secure Cortex Agent with proper RBAC, zero-downtime updates, and cost controls. The running example is the same release notes intelligence agent from Part 1. The Developer Workflow: Personal Database to Production Develop in Personal Database Each developer creates agents in their own Personal Database. No CREATE AGENT privilege on shared schemas required. CREATE OR REPLACE AGENT &#34;USER$JSMITH&#34;.PUBLIC.release_notes_dev COMMENT = &#39;Dev iteration — testing new orchestration instructions&#39; FROM SPECIFICATION $$models: orchestration: claude-sonnet-4-6instructions: response: &#34;Testing revised response format — include release date in bold.&#34; orchestration: &#34;Always use ReleaseAnalyst first, fall back to ReleaseSearch.&#34;tools: - tool_spec: type: &#34;cortex_analyst_text_to_sql&#34; name: &#34;ReleaseAnalyst&#34; - tool_spec: type: &#34;cortex_search&#34; name: &#34;ReleaseSearch&#34;tool_resources: ReleaseAnalyst: semantic_view: &#34;release_intel_db.data.release_notes_model&#34; ReleaseSearch: search_service: &#34;release_intel_db.data.release_search&#34; max_results: 5$$; The Personal Database agent references production data sources (the semantic view and search service) but lives in an isolated schema. No naming conflicts. No shared-schema privilege requirements. Validate with Temporary Agent Before promoting to production, create a temporary agent. Session-scoped — it disappears when the session ends and doesn’t require CREATE AGENT on the target schema. CREATE TEMPORARY AGENT release_notes_validation FROM SPECIFICATION $$models: orchestration: claude-sonnet-4-6orchestration: budget: seconds: 30 tokens: 16000instructions: response: &#34;Testing revised response format — include release date in bold.&#34; orchestration: &#34;Always use ReleaseAnalyst first, fall back to ReleaseSearch.&#34;tools: - tool_spec: type: &#34;cortex_analyst_text_to_sql&#34; name: &#34;ReleaseAnalyst&#34; - tool_spec: type: &#34;cortex_search&#34; name: &#34;ReleaseSearch&#34;tool_resources: ReleaseAnalyst: semantic_view: &#34;release_intel_db.data.release_notes_model&#34; ReleaseSearch: search_service: &#34;release_intel_db.data.release_search&#34; max_results: 5$$;-- ValidateSELECT SNOWFLAKE.CORTEX.DATA_AGENT_RUN( &#39;release_notes_validation&#39;, $${&#34;messages&#34;: [{&#34;role&#34;: &#34;user&#34;, &#34;content&#34;: [{&#34;type&#34;: &#34;text&#34;, &#34;text&#34;: &#34;What features reached GA on September 16, 2026?&#34;}]}]}$$); Temporary agent limitations: no versioning (COMMIT, aliases), cannot convert to permanent, and DATA_AGENT_RUN requires a fully qualified name for permanent agents (temp agents use the session namespace). Promote with COPY GRANTS The production agent already has grants to release_agent_consumer and analyst_role. COPY GRANTS preserves all of them during the replacement. USE ROLE release_agent_owner;CREATE OR REPLACE SECURE AGENT release_intel_db.agents.release_notes_agent COMMENT = &#39;Production release notes agent — v2, revised orchestration.&#39; PROFILE = &#39;{&#34;display_name&#34;: &#34;Release Intel&#34;, &#34;avatar&#34;: &#34;release-notes.png&#34;, &#34;color&#34;: &#34;blue&#34;}&#39; COPY GRANTS FROM SPECIFICATION $$models: orchestration: claude-sonnet-4-6orchestration: capabilities: analytical_search: true tool_not_accessible: accept budget: seconds: 30 tokens: 16000instructions: response: &#62; You are a release notes intelligence assistant. Include the release date in bold for every feature mentioned. Be precise about GA vs Preview status. orchestration: &#62; Always use ReleaseAnalyst first. Fall back to ReleaseSearch only when ReleaseAnalyst cannot answer.tools: - tool_spec: type: &#34;cortex_analyst_text_to_sql&#34; name: &#34;ReleaseAnalyst&#34; - tool_spec: type: &#34;cortex_search&#34; name: &#34;ReleaseSearch&#34;tool_resources: ReleaseAnalyst: semantic_view: &#34;release_intel_db.data.release_notes_model&#34; execution_environment: type: &#34;warehouse&#34; warehouse: &#34;COMPUTE_WH&#34; ReleaseSearch: search_service: &#34;release_intel_db.data.release_search&#34; max_results: 5$$;-- Verify grants survivedSHOW GRANTS ON AGENT release_intel_db.agents.release_notes_agent; The operation is atomic. Consumers experience no interruption. Note: COPY GRANTS must come after PROFILE and before FROM SPECIFICATION — the clause ordering matters. RBAC Design -- Owner role: full specification access, deployment authorityCREATE ROLE IF NOT EXISTS release_agent_owner;GRANT CREATE AGENT ON SCHEMA release_intel_db.agents TO ROLE release_agent_owner;GRANT USAGE ON DATABASE release_intel_db TO ROLE release_agent_owner;GRANT USAGE ON SCHEMA release_intel_db.agents TO ROLE release_agent_owner;GRANT USAGE ON SCHEMA release_intel_db.data TO ROLE release_agent_owner;-- Consumer role: invocation only, no spec visibilityCREATE ROLE IF NOT EXISTS release_agent_consumer;GRANT USAGE ON DATABASE release_intel_db TO ROLE release_agent_consumer;GRANT USAGE ON SCHEMA release_intel_db.agents TO ROLE release_agent_consumer;GRANT USAGE ON AGENT release_intel_db.agents.release_notes_agent TO ROLE release_agent_consumer;-- Transfer ownership away from ACCOUNTADMIN (see Part 1)GRANT OWNERSHIP ON AGENT release_intel_db.agents.release_notes_agent TO ROLE release_agent_owner REVOKE CURRENT GRANTS;GRANT USAGE ON AGENT release_intel_db.agents.release_notes_agent TO ROLE release_agent_consumer; Three roles, four capabilities, one asymmetry: everyone can invoke, few can read. The consumer role has exactly 4 grants: USAGE on database, schema, agent, and warehouse. Zero grants on the DATA schema. Zero SELECT on any table. Zero GRANT_OPTION anywhere. Graceful Degradation: What Happens When a Tool Goes Down With tool_not_accessible: accept, the agent continues when a configured tool is unavailable. Here is what we observed with the Cortex Search service suspended: Three query shapes, three routing paths — only one ever needs a retry. The failover was genuinely graceful. Structured queries were unaffected. Conceptual queries took ~11 seconds longer (retry loop) but returned correct answers via SQL fallback. The three tool_not_accessible values: Three settings, two real policies: availability or compliance. Cost Optimization The deployment has three cost levers: The Cortex Search service was the largest controllable cost. With TARGET_LAG = &#39;1 hour&#39;, it refreshes hourly regardless of data changes. For a dataset that updates weekly, that&#39;s waste. -- Suspend to stop refresh costsALTER CORTEX SEARCH SERVICE release_intel_db.data.release_search SUSPEND;-- Resume when needed (rebuilds from checkpoint)ALTER CORTEX SEARCH SERVICE release_intel_db.data.release_search RESUME; Break-even analysis: The search service costs ~0.04 credits/day. Each agent invocation costs ~0.0034 credits. At fewer than 12 conceptual queries per day, the search service costs more than the queries it serves. Recommended suspend threshold: fewer than 5 conceptual queries/day. Both settings scale together: low volume suspends, high volume refreshes hourly. Trade-Offs When to Use Secure Agents The specification contains proprietary orchestration logic The agent is shared with external accounts or through Native Apps Multiple teams have USAGE but should not see tool configurations When to Avoid Internal agents where all consumers are trusted Development environments where spec transparency aids debugging What to Monitor -- Invocation volume and errorsSELECT execution_status, COUNT(*) AS cntFROM SNOWFLAKE.ACCOUNT_USAGE.QUERY_HISTORYWHERE query_text ILIKE &#39;%DATA_AGENT_RUN%release_notes_agent%&#39; AND start_time &#62; DATEADD(&#39;day&#39;, -7, CURRENT_TIMESTAMP())GROUP BY 1;-- Ownership hasn&#39;t reverted after replacementSHOW AGENTS LIKE &#39;release_notes_agent&#39; IN SCHEMA release_intel_db.agents;-- Check: is_secure = true, owner = RELEASE_AGENT_OWNER (not ACCOUNTADMIN) Key Takeaway The deployment pipeline is now complete. Personal Database → temporary agent → secure production agent with COPY GRANTS. Before September 2026, each step had [&#8230;]]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*rtkj2uBAqpIQV-WAzVJadw.png" medium="image"></media:content>
            	</item>
		<item>
		<title>Coding an Agent: Steering a Local Model</title>
		<link>https://towardsai.com/p/machine-learning/coding-an-agent-steering-a-local-model</link>
		
		<dc:creator><![CDATA[Enzo Lombardi]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:30:36 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/coding-an-agent-steering-a-local-model</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Enzo Lombardi Originally published on Towards AI. Directional Edits in DS4, or why a slider can do what a fork of the weights cannot Most of what you change about a language model’s behaviour you change from the outside. You write a system prompt, you pick a sampling temperature, you ask for brevity in the last line of the message. The model reads those words like everything else and decides how much to care. There is a second lever that most people never touch because most runtimes do not expose it: you can reach inside the forward pass and edit the activations directly, one layer at a time, while the model is running. The article explains DS4 activation steering—using per-layer direction vectors and a scale during inference—to shape model behavior without fine-tuning or loading new weights, including the mechanics of how projections are applied in attention/FFN outputs and what trade-offs (like decode speed) come with it. It then walks through how a steering direction is built from contrasting prompt lists (e.g., succinct vs verbose), demonstrates the effect with an agent-harnessed example measuring verbosity changes, and shows that results depend strongly on context: system prompts shift the baseline, thinking changes the multiplier behavior, and steering competes with the model’s internal deliberation. The author compares steering to abliterated weights (a permanent weight edit) using a refusal/compliance case study, finding steering can match ablation in some settings but is sensitive and context-dependent. Finally, it covers how to start steering after the prompt (“from user”) to avoid altering how the model interprets the scene, outlines measurement limits and failure modes, and concludes that steering is more like a calibrated instrument than a one-time fork of weights—requiring re-sweeping across the conditions you actually intend to run. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*jOao6kz_lfj1-OR-fPe18A.png" medium="image"></media:content>
            	</item>
		<item>
		<title>Machine Learning Basics 5 Things I Wish I Knew First</title>
		<link>https://towardsai.com/p/machine-learning/machine-learning-basics-5-things-i-wish-i-knew-first</link>
		
		<dc:creator><![CDATA[Programming India]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:30:26 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/machine-learning-basics-5-things-i-wish-i-knew-first</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Programming India Originally published on Towards AI. Five simple ideas that make ML tutorials, code and errors finally make sense The first time my model scored 98% accuracy, I took a screenshot. I felt like a genius for about ten minutes. After the opening anecdote, the article lays out five beginner-friendly “machine learning basics” that explain why tutorials can fail in practice: start with data quality rather than algorithms, understand that models only learn from the features you provide, always split training and test data to avoid misleading scores (including guarding against data leakage), interpret performance using the intuition of overfitting vs. underfitting, and remember that ML models predict patterns without true understanding—so wrong outputs typically trace back to bias, missing features, evaluation mistakes, or model capacity. It wraps up with a practical project checklist, an FAQ, and a short conclusion encouraging readers to learn by experimenting on small datasets and intentionally breaking/fixing assumptions. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*LGMZ8mtXr-EnLMjluVKR7g.jpeg" medium="image"></media:content>
            	</item>
		<item>
		<title>The Fundamental Theorem of Calculus, Explained Like a Friend Would</title>
		<link>https://towardsai.com/p/machine-learning/the-fundamental-theorem-of-calculus-explained-like-a-friend-would</link>
		
		<dc:creator><![CDATA[Kamrun Nahar]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:27:36 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/the-fundamental-theorem-of-calculus-explained-like-a-friend-would</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Kamrun Nahar Originally published on Towards AI. See why the area grows at the curve’s height, why plus C cancels, and where it finally breaks. Next time you’re in a car, look at the dashboard. There are two numbers there that most people never think about together. Watch the rectangles shrink until the area is exact, which is the whole idea of an integral in one loop.After the dashboard hook, the article develops calculus as “area in motion”: starting with Riemann sums (rectangles approximating a curved region), it explains why different sampling rules converge toward a limit (the definite integral), and how the speedometer/odometer analogy maps rate to accumulated total. It then ties differentiation and integration together via the area-so-far function A(x): its slope matches the original curve, which is the core statement of the Fundamental Theorem of Calculus. Along the way, it corrects common myths (integrals aren’t just plug-in formulas, signed area can be negative, shortcuts fail when the function blows up or isn’t continuous, and many real integrals arise from measurement rather than closed forms). Finally, it broadens the intuition with applications—electricity bills, medicine’s AUC, machine learning’s AUC/ROC area, probability distributions—and concludes by showing why the “+ C” constant cancels in definite integrals and how the whole story explains what you’re already seeing on a trip meter. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:640/1*zSTAL4Dg9FVPXabHW-UNUg.gif" medium="image"></media:content>
            	</item>
		<item>
		<title>Deepseek-V3: Multi-Token Prediction — Part 3</title>
		<link>https://towardsai.com/p/machine-learning/deepseek-v3-multi-token-prediction-part-3</link>
		
		<dc:creator><![CDATA[Prachi rise]]></dc:creator>
		<pubDate>Tue, 06 Oct 2026 07:26:33 +0000</pubDate>
				<category><![CDATA[Latest]]></category>
		<category><![CDATA[Machine Learning]]></category>
		<category><![CDATA[Towards AI - Medium]]></category>
		<guid isPermaLink="false">https://towardsai.com/p/artificial-intelligence/deepseek-v3-multi-token-prediction-part-3</guid>

					<description><![CDATA[Last Updated on October 6, 2026 by Editorial Team Author(s): Prachi rise Originally published on Towards AI. Deepseek-V3: Multi-Token Prediction — Part 3 This is the full series of Deepseek-V3 technical report, where i explain all the technical details in simpler words with code implementation and explanation. Deepseek-v3 MTP(Multi-token prediction)The article introduces multi-token prediction (MTP) as an extension of conventional next-token autoregressive modeling, explaining how training with multiple future tokens provides stronger learning signals. It then describes DeepSeek-V3’s approach using sequential MTP modules, where predictions at one depth feed into the next depth. The piece breaks down the internal MTP components (shared embedding, shared output head, transformer block, and a projection matrix), outlines how hidden states and future token embeddings are combined through normalization and projection, and shows the step-by-step flow of MTP depth k (including selecting future token embeddings, updating hidden states via transformer blocks, and producing logits with the shared head). Finally, it provides a PyTorch-style code implementation and concludes with the key takeaways: MTP improves representation learning for future-token prediction, DeepSeek-V3 stacks sequential MTP modules, and the method can also be leveraged for speculative decoding during inference. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI]]></description>
		
		
		
		<media:content url="https://miro.medium.com/v2/resize:fit:700/1*RYrxQtN2ItPVxxh12e2WJA.png" medium="image"></media:content>
            	</item>
	</channel>
</rss>
