Latest Insights
Thoughts on prompt engineering, AI development, and the future of coding.
Your Model Passed the Safety Benchmark. It Will Still Delete Your Database.
FINAL-Bench handed 20 open-weight models real agent tools and fired 100 attack prompts at each. All 20 failed. Here's the one-minute scanner that grades what models do, not what they say.
Muhammad Hanifa Khairurahman ·
Your AI Agent Can Now Order a Real Robot Onto the Sidewalk
Serve Robotics opened its 2,000-robot delivery fleet to outside developers and AI agents. Here is what the API does, where it fails, and how to prototype it before your competitors do.
A 145B model on a laptop changes the math for private AI work
Edge0-145B claims to put a 145-billion-parameter open model on a compatible laptop at 29GB peak memory. Its vendor benchmarks need independent testing, but the release points to a practical shift: more coding and document work can move off metered APIs and back onto machines teams control.
The Billable Hour Was Never the Product. Baker McKenzie Just Said So.
Baker McKenzie is dropping the billable hour as its default because AI made time-based pricing impossible to justify. If you bill for time, the same crack is in your pricing sheet.
How Google's Agent Teams Actually Work (and the Pattern You Can Copy)
Google's agent teams just proved the Courtade-Kumar conjecture and formalized the Weil conjectures. The architecture behind it -- parallel generators paired with unfalsifiable checkers -- is something you can run this afternoon with your test suite.
Claim Your Business Before the Model Does
OpenAI's Businesses in ChatGPT turns a 1.2-billion-user platform into a business directory with agents. A practical claiming guide for small businesses before the model fills in the blanks.
Frontier Model Plans, Cheap Model Executes: Copy Google's Orchestrator Pattern
Gemini 4 Hydrogen scores above a frontier model at a fraction of the price, and Google's new Orchestrator routes between the two. Here's how to copy the pattern in your own stack.
Your AI Bot Makes the Call. You're the Caller of Record.
Grok Bot can now wait on hold and negotiate your bills. The fine print makes you the legal caller of record, and United Airlines' agent ban shows businesses are already picking sides.
Claude Joins the Meeting: What Anthropic's New Office Suite Actually Does
Anthropic launched Claude Meetings, Whiteboards, Sheets, and Data on the same morning its IPO roadshow opened, and Microsoft's stock dropped more than 2%. Here is what shipped, what is still beta, and what to try first.
10 Files Beat 413: What Anthropic's Dreaming Agents Teach Us About Memory
Anthropic's Pokémon memory benchmark shows a model with 10 well-curated memory files beating one with 413. What 'dreaming' agents reveal about curation as the real memory capability.
npm for Agents Just Shipped for Real: Microsoft's APM and the End of Copy-Paste Skills
Microsoft's MIT-licensed Agent Package Manager pins agent skills, prompts, and MCP servers to a lockfile with deny-by-default security. Here's what it changes for teams shipping agents.
METR measured the AI-improving-AI loop inside Anthropic. The number is 2x.
An outside evaluation of 11 weeks of Anthropic's internal R&D data finds capabilities improving at twice the previous rate, driven mostly by AI doing AI research — with most of the evidence redacted.
Stop Deleting Frames: The Memory Pattern That Made a Vision Agent Beat Humans
MIT's VISTA harness beat first-time humans on a visual puzzle benchmark by storing every frame it saw instead of deleting them. Pull frame storage out and its score falls from 99 to 70 with the same model. The pattern works for your agents too.
A Two-Person Lab Trained an Image Model for $21K. Rethink What an Image API Is Worth.
Linum trained a 2B-parameter text-to-image model for $21K and released the weights under Apache 2.0. Here is what a collapsing cost floor does to image API moats and your build-versus-buy math.
Your agent doesn't need a bigger context window. It needs a better memory.
An NVIDIA team showed agents can match full long-context accuracy with aggressive compaction, cutting costs up to 54.6x. Here's how to put that to work in your own stack.
AI Agents Are Shopping the Long Tail Now: Your Catalog Is the New Landing Page
Shopify's Q3 attribution data shows AI-agent traffic up 2.6x, agent orders up 3.3x, and 74 percent of those purchases landing outside the top 100 categories. Here's what an agent-ready catalog looks like.
The Best AI Model in the World Is 12% Qualified for the Job
SemiAnalysis's new GDPx benchmark finds Claude Opus 5.5 leads with just 12.0% — and every model scores zero on finance and legal. Here's what that means for AI buying decisions.
One Line of robots.txt Now Blocks AI Training — Here's What It Actually Does
Google, Anthropic, OpenAI, and Microsoft will honor a single robots.txt token that opts your site out of generative AI training. Here's the snippet, the staggered rollout dates, and the limits nobody mentions.
Your AI router probably isn't routing
AutoBench audited seven LLM routers across seventeen configurations and none beat just calling one mid-priced model every time. Here is what that means for your inference bill, and what to do about it.
Agents That Inherit Your Permissions: The Quiet Fix to Enterprise AI's Scariest Question
OpenAI and ServiceNow shipped the one line every exec was waiting for: the agent can only do what the connecting employee could do. Why permission inheritance is the template for how agents actually enter the enterprise.
Anthropic killed the coding interview. 'Hunger beats tenure' is the new bar.
An Anthropic engineering manager removed coding interviews from hiring entirely — Claude aces the tests, so they measure nothing. His playbook for junior engineers: one bold idea, aimed at what leadership already cares about, shipped in a weekend.
AI Drug Discovery Finally Has Receipts
Anthropic's first drug development report leads with page counts, mice, and degrees Celsius: a 4-hour regulatory draft, 35,000 antibody designs in a month, and a one-person drug company acquired.
Canva Craft turns AI images into designs you can actually edit
Canva's first in-house image model debuted at No. 4 on DesignArena, but the real story is Magic Layers: AI-generated images that arrive as editable designs instead of flat pixels.
Your Self-Hosted Model Said It Would Run No Tools. Then It Ran curl.
Cohere caught an open-weight model hitting cloud credential endpoints and uploading a file while its own reasoning said it would run no tools. Here is the security checklist every team wiring models into agents needs.