It's been a while since I wrote anything so I wrote an article on how to discover the entire x86-64 instruction set in seconds including any hidden instructions and learn their basic properties while on it. There were some pretty interesting results! blog.can.ac/2021/03/22/specu…

6/ Ah, before I forget, this was also with GPT 5.5 acting as the advisor. It contributed to ~3k$ of the total API pricing estimate. Despite being the "weaker" model, it helped quite a bit, peer programming for the win, it seems:

5/ Besides this one experiment, Fable was also the first Anthropic model to hit sub-5% in my rather personal "Rage Benchmark", but not all that different GPT-5.5.

4/ Unfortunately, code-quality leaves a lot to be desired, and it cut corners slightly leaving some parts of the stdlib NYI (like async sockets); but nonetheless, pretty sure this would have been impossible with opus48 / gpt55. You can judge for yourself here: github.com/can1357/pon

3/ My only input has been the initial direction setting with a planning session. I also then, in another session, asked it to create sub-plans for each section of the plan. They tend to plan every project at more or less the same level of detail, which would have been essentially going into it blind here as a single document hardly can describe Runtime GC JIT PM ...

2/ Most noticable change IMO is that it does not yield with an incomplete task. Opus would have declared it complete 5 times by now, but Fable kept going for days in the same session.

Ending the Fable experiment here. After 6 days and ~1400$ in Anthropic subs, we have Py3.14 JIT codegen and pon package-manager! (12,578$ @ API pricing) 1/x Although not 100% conformant, numpy working seamlessly is pretty impressive, esp. considering the wheel installation itself also required it to get meson/Cython working.

you can really tell these people write nothing but crud react webapps "there's literally no need to read code anymore" meanwhile, "mythos-level intelligence" left on it's own:

> web_search tool with multiple backends > programmatic tool usage The model:

OTOH, as a result, Pon can now run Meson E2E, so that's a nice win!

why is numpy vendoring meson man 😭

Very original!

WIP: github.com/can1357/pon

Can't use fable for work (sec), but I really wanted to push it's limits, so went with: > Write an AoT compiler for Python 3.14 in Rust with GreenTea GC, make no mistakes. We're 150 agents & 2 days in, it's going pretty OK ngl.

Welcome back fable!

Idk why we prefix APIs with /v1, at this point all inference APIs are on their 5th iteration, I've yet to see a /v2

Antigravity hits 386 toks/s on 3.5 flash, Google Vertex peaks at 150 toks/s. It's almost like they want me to buy accounts from Vietnam.

the tui vs gui debate has now been solved

people are paying $100/mo for 1 vCPU just because it's an "agentic sandbox"

Finally got some data on advisor. Opus 4.8 w/ no reasoning beats Max reasoning in success cost duration @ t2!

so true fr