5.6 Sol (medium) in Codex read it in 2 minutes, and only because it didn't know how it was hidden beforehand. The agent can just save the script to reuse the next time. I hope people properly test such extraordinary claims before posting this.
To be honest, I'm also leaning this way, especially because of the hardcore anti-AI stance, so much that Zig will close security vulnerability issues on Codeberg if you mention that they were found with LLMs. I don't think that this is a good approach.
You can't do "redesigning the existing Zig code in a way that eliminates not only the current bugs but also prevents similar ones from happening in the future" without actually changing Zig itself.
It feels like the first half of blog post is less of "thoughts on the Bun Rust Rewrite" and more "I don't like Jarred, he's a bad programmer and manager".
Maybe I'm wrong, but it strongly feels this way. I'm not saying that Andrew is right or wrong, it's just that you could throw out most of the first half of the post and not lose anything actually on topic.
> But having graduated from the Thiel Fellowship school of thought rather than university, he was essentially groomed from a young age into uncritically embracing the Silicon Valley mindset, and he took venture capital.
> Jarred was a stinky manager. Poor communication, unrealistic expectations, low empathy, no experience. Just a total shit show, from an employment perspective.
> Jarred was already writing slop well before he had access to LLMs
Sonnet 5 is a huge token hog, though, it uses far more reasoning tokens than Opus models while being priced at $2/$10 with promo, and $3/$15 (usual Sonnet price) afterwards.
It seems to be extremely economical - 4x better reasoning efficiency compared to Opus while being priced at $2/$6. For comparison, GPT 5.4 is $2.5/$15, GPT 5.5/5.6 are $5/$30, Opus 4.8 is $5/$25, Fable is $10/$50.
Seems to be found as a part of Patch The Planet [0] which is basically OpenAI giving model access and Trail of Bits using them to find vulnerabilities in OSS projects.
Honestly it's a bit of a shame. I checked and they could've shortened their base64 payload by 304 chars by removing all comments except the top two congratulatory ones, or by 524 if they removed those too.
OCRing this is a nightmare and is a good benchmark to any self-proclaimed good OCR/vision model.
I think though it could likely be easily OCR'd if you give the image to any decent agentic harness with a good vision model, e.g. newest Claude/GPT ones, and tell them to split the image per lines, and then just OCR each line individually.
I wonder if the script itself was written by an LLM before obfuscation? There seem to be a lot of comments in it, but in this case it's still ok :)
I think this was true with older models, but at least with GPT 5.5 it can genuinely tell you "no issues found" after a few passes of finding real issues.
The title cost is only if this was raw API usage, but it was included in a subscription, so it's a small subset of the $200 plan:
> I upgraded to the Claude Max $200/month plan (I was previously on $100/month) to increase my Fable allowance for the remaining time until the July 7th Fablepocalypse, when even Claude Max subscribers will have to pay full API cost for the model.
I really wonder if Anthropic will stick with their decision to keep Fable on extra usage credits until they "get more compute", especially in the light of GPT 5.6 very likely coming out next week (it's confirmed to have the exact same pricing as GPT 5.5)
Not yet, unfortunately, but I might in the future. To be honest, it's nothing unique. I got inspired by https://github.com/allthingsida/idasql which I initially used, but it had a lot of bugs, a big codebase size, and IDA's C++ API is really easy to misuse and corrupt a DB, so I had GPT 5.4/5.5 make a new one for itself based on IDA's Python SDK, which is official and doesn't need weird SQL hacks.
Then recently I found https://github.com/bkerler/ida_rpc which seems to be ~60% the same thing as the one I have, the only big difference is that I do not give any special commands to LLMs, they just have to write Python in scripts/inline heredocs to interact with IDA. This lets them do a lot more interesting things since they get a full programming language.
This is an example of how LLMs work with idagent (`ida` is implicitly imported, ida.types, ida.comments is helper's own wrappers): https://paste.debian.net/hidden/cf46a122
More interesting example that was used to let the LLM/me track the rename progress for the initial function renames + gaps (code-looking like bytes that weren't inside of functions, IDA's autoanalysis missed some real functions). Although the game turned out to be small enough with only ~1500 real game functions that needed renames, which was done in ~10 hours of agent time total I think (I didn't parallelize with multiple agents).
https://paste.debian.net/hidden/bf458b3a
To be honest, you can probably have an agent vibecode a similar MVP tool to the one I have in about an hour-two :)