1081
DeepSeek v4.1 Flash Is Now Our Best Hacking Model
DeepSeek’s 11/11 result showed why advanced agent benchmarks need to check both the outcome and the attack path: our audit confirmed six planned exploits and found five unexpected routes.
1082
When Genius Fails: The Intellectual Arrogance of the AI Labs
From Situational Awareness’s Blow-up to Materials Science to the HuggingFace Hack
1083
747s and coding agents
Comparing a career flying the 747 and a career coding with AI assistance.
1084
What would a serious AI product look like?
Every “AI” tool is missing critical features that you would need if you wanted to do real work with them. What would those look like, if they existed?
1085
A primer on GPUs and kernel optimization
The basics of GPUs, profilers, distributed computation, and kernel optimization
1086
Are LLM merge rates not getting better?
I was reading the metr article on how llm code passes test much more often than it is of mergeable quality . They look at the performance of llm s doing programming when the success criterion is “passes all tests” and compare it to when the success criterion is “would get approved by the maintainer”. Unsurprisingly…
1087
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
A working tmux setup for running Claude Code and other AI agents: session-per-project naming, pane layouts, send-keys for scripting an agent, capture-pane for reading one back, and the status-line monitor. Plus the limits you hit at four agents.
1088
Using LLMs to trace alchemical knowledge and decode 17th century letters
Using GPT-6 and Opus 5.5 to trace alchemical knowledge and decode 17th century letters
1089
Designing more efficient quantum circuits with Claude
Working largely on its own, Claude designed quantum circuits that improve on the best published versions, in some cases by an order of magnitude or more.
1090
1091
We made Grok 4.5, GPT-5.5, and Claude build the same apps
Grok 4.5 just launched. So we had it, GPT-5.5, Claude Opus 4.8, and Fable 5 one-shot the same interactive apps, then measured latency and cost. Here is who won.
1092
Introducing LFM3: Every Modality for Specialized Intelligence
LFM3 is Liquid AI’s new open-weight model family: 2B-class text, vision, audio, omni, and decision models built to fine-tune and run fast on any device.
1093
How AI is affecting productivity and jobs in Europe
Artificial intelligence promises to reshape economies worldwide, but firm-level evidence on its effects in Europe remains scarce. This column uses survey data to examine how AI adoption affects productivity and employment across more than 12,000 European firms. The authors find that AI adoption increases labour…
1094
Fixing Tool-Call Accuracy in vLLM
Tool call accuracy is an important part of serving open-weight AI models. When your tool calls aren’t accurate, the model builds up on nonsensical context. We’ve been carrying our own fork of vLLM to hit 100% in tool call schema accuracy with Kimi and DeepSeek. Fortunately, vLLM 0.31 added support for constrained…
1095
Matrix Multiplications on GPUs Run Faster When Given “Predictable” Data (2024)
Great minds discuss flops per watt.
1096
Just say no to OpenAI and Anthropic
I recently read End Times Fascism and the Fight for the Living World by Naomi Klein and Astra Taylor. It's an excellent book and I highly recommend that everyone read it. Tech oligarchs make up an important component of the end-times-fascist coalition. One thing that stood out
1097
Cf: The Agentic CLI for the Cloudflare API
We are releasing cf, our new command-line tool that mirrors the entire Cloudflare API and supports programmatic TypeScript configuration. We are also open-sourcing Forge, our internal SDK generator.
1098
Hamilton-Jacobi-Bellman Equation: Reinforcement Learning and Diffusion Models
Why the HJB is Bellman's equation in continuous time, why continuous time matters, and how to solve the resulting control problem with neural policy iteration.
1099
How we made a text-to-speech model respond in sub-50 ms
10 RPS with p95 TTFA under 50 ms on a single H100: how we optimized Qwen3-TTS serving.
1100
How does Shazam work?
Explore how Shazam and song identification work through interactive visualizations: spectrograms, constellation maps, hash fingerprints, and time-offset matching.
1101
1102
It wasn't AI that flipped Proton's charge
Proton's CEO called a request for security testing "loser thinking". Using AI to write code is fine, but for a privacy company, the guardrails around that code are the product.
1103
RTK reports token savings, but our cost benchmarks disagree
We tested RTK (Rust Token Killer) with Claude Code on Fable 5.0, and OpenCode with DeepSeek V4 Pro 0813 on Terminal-Bench 2.1.
1104
Policy on the AI Exponential
In one of the side plots to The Lord of the Rings , two of the Hobbits attempt to rouse Treebeard—a wise but ponderous sentient tree—to defend his forest from an army that is cutting it down. The problem is that Treebeard operates at a very different speed than the Hobbits. It takes him a full day simply to say hello…
1105
Learnings from paying artists royalties for AI-generated art
A retrospective on Tess.Design, our attempt to make an ethical, artist-friendly AI marketplace. We launched Tess in May 2024 and shut it down in January 2026.
1106
Claude Managed Agents
Claude Managed Agents is Anthropic's suite of composable APIs for building and deploying cloud-hosted agents at scale. See what's new and how to get started.
1107
Redeploying Fable 5
Anthropic is redeploying Claude Fable 5 starting July 1, with updated cyber safeguards and a new jailbreak framework.
1108
Pacing model development in an era of cyber-critical capabilities
OpenAI is strengthening monitoring, alignment, and security for frontier AI models. See how new safeguards are guiding the pace of model development.
1109
I don't know how you get here from “predict the next word”
I recently tried refine, an AI tool for refining academic articles, developed by Yann Calvó López and Ben Golub.
1110
Every SaaS business will become a harness around a model
Why SaaS business is now a harness one.