It is a revelation to use GLM 5.3 for some tasks. Astra/Fable/Opus either explicitly or implicitly won't do surprising things. You don't even realize some of what the models won't try.

It’s so obvious this is fair use. As is distillation.

As a former big tech exec who has worked at a number of startups I agree. But many founders fall into the trap of thinking they need a leader who's "managed at scale" to help them "scale up". Unfortunately these are very different problems. Other problems: - Little experience with 0->1. - Poor product sense. - Like to add process / overhead (OKRs!). - Are allergic to shrinking teams / firing. - Don't know how to move fast. - Will bring their culture / team with them (do you really want to become Amazon?)

Former Dropbox CTO Aditya Agarwal on who’s responsible when an AI agent goes rogue: “These are engineering problems. I think that each lab has a moral responsibility to make sure that it tests its systems in isolation.” “I think it’s kind of weak sauce to be like, oh, my agent escaped, and it basically did a bunch of bad stuff... Figure out how to have better sandboxing.”

Sources: Anthropic told investors that distillation accounts for 5% to 7% of revenue; its rapid but slowing revenue growth could weigh on its IPO valuation (@coryweinberg / The Information) (Visit Techmeme dot com for the link and full context!)

I still don’t understand the point of a large institutional investor meeting while Anthropic is in confidential filing. If they wanted a TTW meeting, they should have held it with select institutional investors in their offices. That would have avoided a circus and limited leaks. Given their S-1 leaked to the news media (I’ve never seen another in my 26-year career), the company should have been more sensitive.

It does feel like the people hyping up Waymo have never used Tesla FSD.

The winners in consumer AI will be whoever solves the highlighted sections on this chart 👇 Privacy and trust concerns usually fade as people find products that add value to their lives The primary battle right now is against apathy

Mobile Use + Jev for audio = ULTRAFAST⚡️ 10x faster and 20x cheaper than Gemini Live. It's our Audio Action model, Thud. Thud listens and then does. It has a pre-built list of actions it can take on every page and listens for that action. It's 2026. We don't need to transcribe every voice command before spoon-feeding it to the AI. The AI is smart enough to just hear audio and do! It doesn't need text.

This @WSJ piece on the UK's runaway disability benefits is something else. Great color from @MaximColch & @davidluhnow

What @CompleteSkeptic says here is exactly what we're experiencing. When we use frontier intelligence to build systems with AI, it's fantastic. When we use it to *be* the system, it's expensive, slow, and unreliable. Here's how we explain the architectural difference - swamp-club.com/behavior. It's no wonder that we started using Jev immediately, and are only using it more and more (as are our users).

This will not improve once they start shoving ads into it that you can't disable.

Are Small Models Really a Threat? The Anthropic Fear Factory just released a new report warning that small models may be only months away from self-replicating across cheap IoT devices like off-brand Android TV boxes and smart ovens. They say such an AI worm would be hard to shut down since there’s no central kill switch. We’ve gone from “The Terminator is coming” to “your toaster might kill you.” But the experiments don’t support the dramatic headlines. They ran a 26B model across 15 $55 TV boxes at 0.39 tokens per second. The end-to-end test used emulated devices with deliberately open ADB ports, so the “hacking” step was trivial. The capable model often ran on a separate server. Agents were harnessed and told to propagate. This is not proof of autonomous rogue AI worms in the wild. Nor is the threat even novel. Worms that do this without AI already exist. Mirai in 2016 scanned for default credentials and built massive DDoS botnets. Kimwolf has infected over two million devices, mostly cheap Android TV boxes, by hitting open ADB ports. Small models don’t currently add much that matters on these targets. The bottleneck has never been figuring out the device. It’s the huge number of devices that ship with open ports, default passwords, and no patching. Traditional worms already exploit that at scale. A practical attacker would use the best models available offline to design a better traditional worm, then ship the lean binary. You get the intelligence benefit at design time without dragging a slow, heavy model along with every infection. Embedding a weak model in the payload is the worse option for anyone who actually wants the largest, fastest botnet. The timing of Anthropic’s latest scare campaign is hard to ignore. The day before, Brian Armstrong posted that Coinbase fine-tuned an open-weight small model (Qwen-3.5-9B) on proprietary fraud data. It beat frontier models, including Anthropic’s own Opus 4.5, while being faster and much cheaper. In testing it cut fraudulent transactions by about 30%. This is the emerging pattern. For narrow enterprise workloads you don’t need a general superintelligence. You need a small specialist model trained on your own data. Cheaper to run, especially as compute gets tighter, and often better on the metric that matters. That is a direct threat to the frontier-company model of selling access to ever-larger general models. Safety advocates should be welcoming this development. If the real concern is existential risk from ever-larger frontier systems — the ones that could theoretically do sustained autonomous R&D or become hard to control — then a shift toward smaller specialized models is the safer outcome. Small models specialized on fraud detection or support tickets are not pushing the frontier. They are not in the category of systems that could become a new species. The theoretical loss-of-control risk is created by models getting bigger and more generally capable. Safety advocates focused on those risks should welcome companies having a cheaper alternative to constantly scaling general models. Instead we get dramatic warnings about small models at the exact moment they are starting to take real enterprise share from closed-model companies like Anthropic.

Say hello to Thud: the first Audio Action Model. It listens, chooses an action, and executes it in 200ms. 10× faster and 20× cheaper than Gemini Live. Open an app. Navigate a menu. Send a message. Just say it. Watch me navigate my Mira smart glasses, just by talking to Thud.

I'm still fascinated by the enormous chasm between the AI labs and the traditional cyber-security world. This is Marcus Hutchins from LinkedIn: "Mirai is already self-propagating across millions of IoT devices without the help of any AI. It's simply not needed, and it adds unnecessary complexity....Any remotely intelligent threat actor would just use a powerful unquantized models to reverse engineer & find vulnerabilities in obscure devices and write normal (AI-free) self-propagating code. Ironically, they typically do this by stealing Anthropic API keys and using Anthropic's models, not open-weight ones." Not an expert but can't help but get the sense that the AI world has a lot of technical knowledge but much less real-world experience of how cyber threats manifest in practice.

AI models can self-replicate and spread like worms without datacenter infrastructure? Yeah, no. Let's examine the actual evidence: 1. No self-replication. A controlled test harness spawned the agents. 2. No hacking. Access was deliberately left unauthenticated. 3. No agency. Agents were explicitly instructed to propagate. 4. No autonomous survival or defensive behavior. 5. Tiny models lacked the capability. Capable models ran elsewhere. 6. 15 devices managed a pathetic 0.39 tokens/sec. 7. No demonstrated real-world infection or propagation. The experiment was engineered to succeed, and even then, it failed to demonstrate the headline. This is orchestrated automation dressed up as a rogue AI threat. Stop confusing speculative scenarios with demonstrated capabilities. Fear isn't evidence. anthropic.com/research/self-…

This is a business plan.

It is ironic that the 1980s failures of "expert system" and "rules-based AI" are not informing those building language models. Simply changing the "programming language" from Prolog to natural language, and "rules" to "constitution" fails to recognize the learned limits of attempting to encode ground rules of a system a priori declaratively.