happened to me too lol. if one of your messages trips anthropic's cyber safety filter, claude code switches you to opus 4.8 and stays there for the rest of the session. i ran ~1k turns on 4.8 in one session without realizing it subagents set to "opus" follow whatever the main session is on, so they get downgraded too your session logs in ~/.claude/projects record which model answered every turn, worth checking after long runs to lock subagents on 5.5, set their model to claude-opus-5-5 instead of just "opus"

Fable 5.1 still feels deeper than Opus 5.5 Next to Fable, Opus lacks depth in thinking and research, so these benchmarks are kind of deceiving a little Fable remains my daily driver. Though one has to say, Opus 5.5 is a huge leap from the previous Opuses, at least you can read it again

an obvious bug in Codex Conversation recaps are made even after you hit the usage limit. Seems like, as conversation summary is just a regular model request, it should fall under the limit. that almost makes me wonder whether one can get convo recaps limitlessly past the cap. Needs fixing?

Grok 4.7's take on Warcraft. 57 minutes of uninterrupted work within one session, 281k tokens spent, no edits. Honestly, I’m impressed

Grok 4.7 is an amazing model. Right now its inference seems kind of slow though, Grok Build works at around 30-50 tokens a second. Wondering whether that’s because of the recent release and if it will get faster.

The question I keep asking myself: how truly good is Jev at generalisation? The obvious leap versus old classifiers is that you can speak to it in broad terms. The catch though is we still don’t have enough honest evals to know how far that leap goes. Published public numbers mostly measure agreement with two frontier models on four vendor workflows, not human ground truth on your own specific distribution. On short, well-specified calls it looks strong. On math, extra hops, noisy state, or a badly written question it gets literal and jagged. So the model is probably better than a frozen classifier and worse than the myth. The only eval that settles it is a few hundred of your own cases, a confidence threshold, and a look at what still needs a human.​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ So eval it, eval it

jev makes me order Chinese over McDonalds and burger chains can you trust jev with your life choices? i think you definitely can

Trained a fly brain to play Flappy Bird - it navigates by vision, dopamine spikes and scent. Flies orient by scent, so the gap smells like food.

in opencode Astra is rushing to end its turns, even with explicit /goal style instructions - in Codex it operates much better. harness truly matters. I guess with frontier Anthropic/OpenAI models you’d just really be better off using native claude code & codex

opencode has just stopped supporting ox-alpha for free what a shame

AI writing useless tests is a context problem, and it's mostly fixable. → The core issue: You cannot trust tests produced in the same context that produced the code. With the implementation in context, the session is conditioned on the very thing it's supposed to check. So the most likely token sequence for the test is a description of what the code does, not what it should do. The result is a self-confirming test that checks nothing. The obvious answer to that is subagents. But if you look at how harnesses actually launch them, turns out that's not really a viable solution. Subagents still inherit the AGENTS.md / CLAUDE.md, the tools, the skills, and a prompt written by the context you're trying to escape. If you look closely at the default prompts harnesses like Claude Code launch their subagents with, it's usually layers of text and tightening definitions. So you contaminate the context of a subagent anyway through the same-harness layer. Not a clean solution. What actually works is graph engineering: a pipeline of minimal, isolated nodes, each getting only the data it needs and nothing more. My setup: • Minimal harness (I use Pi) • CLI invocation • Orchestration logic on top A quick example with e2e test generation and a simple three-step graph: 0. PRD → 1. User Flows → 2. Tests Node 1 sees only the PRD. Its one job is to invent the user flows that will actually happen when this ships, edge cases included. No code and no prior PRD-generation context. The flows get reviewed by a human (or an LLM judge, extra node) and fixed here. Node 2 gets the flows along with the codebase, and writes the corresponding tests. It reads the code, but what to test was already decided upstream by a node that never saw the code. So there's nothing (or at least much less) left to cause self-consistency bias. Then you get tests that check real behavior instead of confirming the code. The name of the game is isolation. Each node should be minimally scoped and blind to the others' reasoning. Want sharper tests? Add more nodes to mirror your architecture, and keep each new node just as blind.

you can tell Ox-Alpha is a Chinese model by the way it outputs its reasoning it streams thinking for 5-7 minutes non-stop even on medium tasks. Gemini family models usually behave differently that gives a strong taste of GLM/DeepSeek (e.g. non-Western) type models