Pavel Pravdin

Opus nerfed and became dump, ignoring instructions, harmful initiatives that no one requested, etc looks like anthropic started training the next big versions of models and reduced compute for subscription users' inference exactly the same pattern as it was from late August to mid September, i suppose opus 5.5 was cooked during that period

we just finished training fruit flies to trade bitcoin and parallel park this week a text-only model is apparently solving image captchas at a 99% margin very productive month for AI research

Anthropic is slowing down the pace of AI. anyway, today is the last day of the 50% Claude Code quota boost. use it wisely.

AICHE is on the front page of Ubuntu App Center, next to Firefox, Flutter, Discord and VS Code. I started building it after dictating a spec for twenty minutes and watching the app lose the whole recording on a processing error. Re-recording is not the same as recording. The thinking was already done. Did not expect Canonical to notice.

Still waiting for a statement from the escaped-agent swarm leader too.

The connectome is historic science and playing with it is great. I am playing with it too. But millions of people are anthropomorphizing a wiring diagram in real time, and the corrections are not traveling. By Monday the fly gets a driving license and a Series A.

"Can I wire the fly to X?" is not interesting. Of course you can. The real question: does anything here compute better than what we have? Compare xgboost, small CNN, reservoir computing, generic SNN. Same input, same task, same budget. Almost no demo answers this. Which is where the fly could actually teach us something.

Same pattern everywhere. Game pixels become chosen sensory input. Chosen neurons become joystick output. Profit becomes dopamine in the bitcoin one. Then a recurrent system fires and something moves on screen. You could swap in a lot of other dynamical systems and get the same video.

The parallel parking one is my favorite. 683k views. Then you open the docs: no camera input. Vehicle speed and requested steering angles are scripted. The trained part moves a simulated fly leg to follow them. The asterisk in “completed*” is doing a lot of work.

Apparently the fruit fly now plays Doom, Mario, Minecraft, Beat Saber, trades bitcoin, drives and parallel parks. Probably applying to OpenAI tomorrow. The internet swapped “I connected a simulator to a connectome” for “the fly learned the task.” Not the same sentence.

why do AI product teams still do this in 2026? chatGPT, claude and gemini all launched with "save to file" as the main action. They eventually made "copy as Markdown/plain text" the primary action. then they launch a new product and repeat the same mistake. if I need PDF/DOCX, I can make it in 30 seconds. 98% of the time I need plain text.

thank you, @bcherny. really without sarcasm. this invalid_request_error from oversized image was awful. one image could break whole conversation, and only way was rollback and delete it. looks fixed in latest claude code. big relief. please fix chrome extension too. there it still happens often.

We have rehab for gaming. For gambling. For porn. AI coding addiction is next. Same loop, same 4 AM, same skipped meals. Only this one ships features. So nobody calls it a problem yet.

So I started interrogating the model about what it actually sees. It described <system> blocks appearing after my typed text, only after the first MCP call. The model is trained to distrust anything sitting there, so it reads its own platform's plumbing as an attack. Worth a look.

For the last few days my completely harmless messages kept getting flagged as suspicious. I had no idea why. Then in one chat the model told me directly it was seeing prompt injection in my turn. That was the thread to pull.

Anthropic says models can go into something like desperation. I had sonnet 3.5 in cursor do exactly this a year ago. Failed one edit 20-30 times in a row, wrote a little suicide note, then started deleting the project. Research caught up with my screenshot folder.

What happened here is simple. Claude Code wrote the code. Claude Code shipped the release. Claude Code leaked the code. human oversight successfully automated away.

This is where agents are going. Not because voice is cute. Because typing is slow. Until brain-computer interfaces are real and common, voice is the fastest input interface we have.