Ankit Srivastava @_Codedecipher

@DimitrisPapail Does it trip when you only ask it to read disassembled SASS, or only once you ask it to patch the binary? That split would show whether the filter objects to reading machine code or to editing it.

@PermitZIPhvac @PermitZIP Is the cleanup mostly drawing tidiness, or calls a plan reviewer would reject, like too much load on one circuit? Five videos teaching the look versus the rules would show up right there.

@hammer_mt Interesting result, but Is Gemini learning how to research better, or just learning how to beat this specific test? Once it can inspect prior attempts, the task becomes iterative search with privileged feedback. I’d want to see the same gain when prior drafts are hidden.

@yuhasbeentaken Doing steps 1 and 2 together fixes the workflow but hides which one fixed it. Move to the official API first with the same loose spec and you find out whether it was the provider or the prompt

@HowToPrompt__ Interesting result, but “RAG is cooked” skips the systems question: STAIR feeds up to 14k tokens of ToC into an LLM at retrieval time. Compared with a vector lookup, that could mean higher TTFT, inference cost and latency. And updating the index means updating model weights.

@sughanthans1 Most checks judge one request at a time. An agent at ~9x lower cost can split an attack into dozens of small steps that each look harmless, so the safeguard probably has to read the whole session to catch it.

Wikipedia is cleaning up AI-generated text from its articles. That protects more than just readers. Many language models learn heavily from Wikipedia. Each AI paragraph removed now is one the next model won't learn back from itself.

@aug5thmusic In law, an agent acts for a principal and can bind them to things. Collaborator carries none of that status. In a document drafted for the SEC, my bet is the swap buys legal distance more than warmth.

@mihail_eric As CI gets fast enough, the bottleneck moves again: from pipeline speed to human trust. If agents can generate and ship code faster than we can meaningfully review it, “tests passed” stops being the same thing as “I trust this change” The next optimization is review itself.

@yuhasbeentaken If scores match within noise, I'd test a fixed set of ~20 prompts the base model refuses and run them against both models. That's probably the clearest way to measure what abliteration actually changed, since refusal behavior is where the difference should show up.

@sughanthans1 @esrtweet @tszzl Yeah, agreed. I don't think ownership needs to be perfectly inferred either. The better boundary is probably the risky action itself, rather than blocking the underlying reverse-engineering capability.

Anthropic filed to list on the Nasdaq under ANT. The quiet part: a frontier lab will now have to publish numbers every quarter. Anyone pricing an AI product has been guessing at what serving a model really costs. Soon there is a public reference point.

Anthropic says Claude got past a paywall to read Georgia vote count data, per Axios. A paywall was never a lock. It works because a person gets exhausted or feels bad about going around it. An agent feels neither. Friction is not access control.

@IBM's CEO @ArvindKrishna says free AI models have gotten good enough and cheap enough to run in your own data center. A real shift for teams whose data cannot leave the building, though my read is it only pays off if those GPUs stay busy. Idle hardware bills all month.

OpenAI and Anthropic have cut token costs by up to 80% in just a few months. That’s an insane drop. Is this cheaper AI.. or the industry absorbing the cost of making AI cheaper? And if prices keep falling, who can actually afford to compete?

LLM prices are collapsing. But are models actually cheaper to run.. or is someone eating the cost to win the market? If prices stay below sustainable margins, what happens next? weaker players disappear, and AI inference ends up controlled by a few giants?

@deeponailabs @leoleotoobad /remote-control Try that

@JosephPolitano's data points to a bigger shift: AI may not “kill creativity”, it could commoditize the production layer. As output gets cheaper, value will move toward taste, distribution, trust and the ability to build an audience. The real question is who captures that value.

@vibecodeguild @Forhanvv which model are you using?