2026
- The 6x Math Jump Looks Like a Loop Fix
TL;DR: An MIT CSAIL team reported that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights and lifted its MATH-500 score from 7.9% to 48.1%, with GSM8K and IFEval rising too. The last column of the …
- Which Agent Platform? Bedrock Managed Agents, AgentCore Harness, or Runtime
TL;DR: Amazon Bedrock Managed Agents, powered by OpenAI, went into preview on September 29, and it is easy to read the launch as a choice between two products: Managed Agents or AgentCore. Read …
- Your Agent Needs to Remember What Passed
TL;DR: Ten pieces, 1% odds each. Guess all ten at once and you need about 10^20 tries. Keep each piece a check confirms and you need about 1,000. In agent loops, the model raises the odds that a …
- The Matryoshka of Humans in the Loop
TL;DR: Amazon Mechanical Turk closed on 30 September 2026. Its own tagline, “artificial artificial intelligence,” is a useful lens for what came after it, because the human-in-the-loop never left the …
- GLM 5.3 on Amazon Bedrock: An Open-Weight Coding Model Behind a Managed API, Tested
TL;DR: GLM 5.3 from Z.ai is a mixture-of-experts coding model with roughly 750B total and about 40B active parameters, a 1M-token context and always-on reasoning, and it is now on Amazon Bedrock for …
- Can You Trust a Swarm's Reasoning?
TL;DR: When a collective of agents coordinates internally and a deployment surfaces only its final broadcast to you (a design I am positing to make the trust question concrete, not something the …
- When Your Benchmark Runs for 88 Hours
TL;DR: On the Dwarkesh podcast, Noam Brown of OpenAI described a system of roughly 10,000 agents that spent 130 billion tokens over 88 hours on a single problem. This is his account, not independently …
- When the Model Becomes an Org
TL;DR: In September 2026, Noam Brown of OpenAI said on a podcast (not independently confirmed) that a system of roughly 10,000 agents spent 130 billion tokens over 88 hours on one of the Millennium …
- Who Pays for the Agentic Web — the Agent or the Advertiser?
TL;DR: Across two series I’ve built both ways a publisher can get paid when the visitor is an AI agent. In one, the agent pays: an x402 micropayment for access, verified at the edge. In the other, the …
- From Prompt to Context to Memory: The Arc, Grounded on AWS
A while back I argued that context engineering had quietly replaced prompt engineering as the skill that decides whether an agent is any good [1]. That piece was the framing. This one is the build. I …
- Who Pays to Un-Bias Your Agent?
TL;DR: If brands will pay to tilt an AI’s answer, someone will pay for the answer that isn’t tilted. That splits the agent economy in two: cheaper, ad-funded agents whose answers lean toward whoever …
- When Ads Bias Your Agent
TL;DR: A sponsored FAQ in agent-read content is a prompt-injection channel with a media kit. To see whether it actually works, I built a small experiment: a fictional publisher serves a clean page and …
- Self-Consistency Is Just Probability Amplification
TL;DR: “Self-consistency”, ask a model the same question several times at a non-zero temperature and take the answer they most agree on (a plurality vote), gives large accuracy gains over greedy …
- Generate-and-Verify: What P vs NP Teaches Us About AI Reasoning
TL;DR: The defining feature of the complexity class NP is polynomial-time verifiability: a candidate answer is cheap to check. The asymmetry that matters shows up in its hardest members, the …
- Stop Guessing Which Coding Agent to Use — Benchmark It on Your Own Tasks
TL;DR: Generic coding-agent leaderboards don’t tell you which model or tool fits your work at a budget you’re willing to spend. So I built a small task set that mirrors my real work and ran an …
- The Quadratic Wall: The Algorithmic Race to Beat O(n²) Attention
TL;DR: Standard self-attention costs on the order of n² in the sequence length n, in both arithmetic and the attention-score memory it materialises. Almost every long-context architecture headline of …
- From Robot Soccer to the Agentic World Cup: Why Games Move AI Forward
🎬 Also available as a blog walkthrough video: a narrated screencast of this post. TL;DR: Twenty years ago, a year-long university project group I was part of built a team of soccer-playing cube robots …
- Why AI Tokens Are So Expensive — and What Actually Makes Them Cheaper
TL;DR: Autoregressive LLM generation has two phases with different bottlenecks: prefill processes input positions in parallel and is usually compute-intensive, while decode produces tokens …
- Vector Search, From the Whiteboard to the Cloud
🎬 Also available as a blog walkthrough video: a narrated screencast of this post. TL;DR: Vector search is easy to picture on a whiteboard: turn text into a list of numbers, then find the stored items …
- Agent Memory Is a Spectrum, Not a Switch
🎬 Also available as a blog walkthrough video: a narrated screencast of this post. TL;DR: “My agent keeps forgetting things” is usually not a model problem. It’s a memory-architecture problem. Agent …
- I Rebuilt a Browser Fact-Checker on AWS, and AgentCore Web Search Was the Missing Piece
🎬 Also available as a blog walkthrough video if you’d rather watch than read. TL;DR: I came across an open-source browser extension that fact-checks videos and live streams in real time, and I wanted …
- Give Every Agent Its Own Computer: AWS Lambda MicroVMs
🎬 Also available as a blog walkthrough video where I walk through the post and the demo. TL;DR: AWS Lambda MicroVMs (launched 22 June 2026) gives every user, job, or AI agent its own isolated, …
- Getting Paid by Agents: A Managed Paywall at the Edge
🎬 Also available as a blog walkthrough video with a narrated tour of the build. TL;DR: In Part 2 I built the buy side: an agent that autonomously pays per article over x402 with Amazon Bedrock …
- Your Voice Clone Is Only as Good as the Reference Clip
🎬 Also available as a blog walkthrough video with narrated diagrams. TL;DR: A voice clone built by conditioning on a reference clip, Qwen3-TTS, in my case, inherits the clip’s cadence, pitch, and …
- MCP Strategies on AWS, Part 2: Tool Design in Code
TL;DR: The AWS Prescriptive Guidance paper on MCP gives the most useful, checkable rules in its tool-design section, so I wrote code to test them. A token-tax counter on a realistic 20-tool GitHub …
- MCP Strategies on AWS, Part 1: Reading the Whole Guide
🎬 Also available as a blog walkthrough video with narrated diagrams. TL;DR: AWS published a Prescriptive Guidance paper, “Model Context Protocol strategies on AWS,” that organizes MCP into three …
- Claude Fable 5 on Bedrock: A Hands-On Comparison, and the Data-Retention Switch You Set First
TL;DR: Claude Fable 5 went GA on Amazon Bedrock yesterday (June 9, 2026), so within a day I ran it head-to-head against Opus 4.8 and Sonnet 4.6 (all three invoked from Frankfurt through EU Geo …
- Why Your Cheapest Model Should Write the Harness
TL;DR: A May 2026 paper separates two capabilities that self-improving agents usually conflate: writing harness updates and benefiting from them. Writing is flat across model tiers: a 9B open model …
- I Built the Agent That Pays — Here's What I Learned
TL;DR: A couple of weeks ago I wrote about HTTP 402 and why AI agents might finally activate the internet’s oldest unused status code. The post sparked a real discussion, so I built it: a research …
- Welcome to the Family: I Sat GPT-5.5 and Claude Opus Down on Bedrock
TL;DR: OpenAI’s GPT-5.5, GPT-5.4, and Codex went GA on Amazon Bedrock on June 1, 2026. To get a feel for it, I wired up two Strands agents, Claude Opus 4.8 and GPT-5.5, and let them chat, with Opus …
- The Service Catalog Pattern for AI Agents
TL;DR: AI agents face the same discovery problem microservices solved with service registries a decade ago, except agents need semantic search, governance workflows, and dynamic capability updates. …
- Your AI Judge Needs a Judge
TL;DR: Most teams I see ship LLM judges without testing them against human labels. The result: judges that are confidently, consistently wrong on 30%+ of cases. Hamel Husain’s “critique shadowing” …
- 8 AWS Guides for Agentic AI — Mapped to the 4 Pillars That Get You to Production
The Gap Between Demo and Deployment TL;DR: AWS released 8 prescriptive guides for building production-ready agentic AI. This post maps each guide to the four pillars that get agents from demo to …
- Intelligence Is About Time, Not Parameters
The Question Every SA Gets Beyond a complexity threshold, larger models become less insightful, the savant regime. “Which model should I use?” I hear it in almost every customer conversation about …
- Hackathon Gamification: A Real-Time Leaderboard You Can Deploy in 5 Minutes
Lisbon, 6 AM, Rising Sun Lisbon at dawn, the best way to start a hackathon day Three days in Lisbon. Not for sightseeing, though the city makes that hard to resist. I’m here for an AI hackathon, teams …
- The Agent Security Stack Nobody Is Building
The Scenario Nobody Planned For It’s 11 PM. Your customer support agent, the AI one, is processing a refund request. It queries the order database, pulls the customer’s payment history, and calls the …
- Building Agents That Read the Web Right
The Other Side of the Coin In a recent article, I made my website AI-agent friendly [1], adding llms.txt, Markdown output, and content negotiation to a Hugo site on AWS. That article was about the …
- Is RAG Still Needed with 1M+ Token Context Windows?
The Kofferklausur, Revisited In September 2024, a colleague asked an audience: “What is RAG?” I answered: Kofferklausur [1]. For non-German speakers: a Kofferklausur is an open-book exam. You bring …
- When Thinking Twice Helps — And When It Doesn't
The Saturday Morning Experiment Last Saturday, I installed a Python library, pointed it at Amazon Bedrock, and asked a model the same questions three times, with zero, one, and three rounds of …
- Your AI Models Have an Expiry Date — A Practical Guide to Model Lifecycle Management
Introduction, The Promise I Made In my previous article [1], I explored the maintenance trap in IT, how software systems are more like plants than stones, requiring constant care. I ended with a …
2024
- How Do You Choose the Foundation Model for Your Generative AI App — Like Your Car?
TL;DR: Nobody buys a car for its engine, and nobody uses a Generative AI app for its model. Users care about functionality, cost and experience. So don’t spend all your design effort picking the …
2023
- Building Your First Generative AI App in Just 6 Minutes — on a Slow Train WiFi
TL;DR: Six minutes on a train with bad WiFi was enough to build a working Generative AI app. PartyRock turned one sentence of description into “MarketGenie”, an app that takes a campaign objective and …