Rasmus Ros

Every "AI makes us dumber" take has been made about calculators and search engines. The panic is half right: tools change what we bother to remember. The new variable is that this one ships inside an engagement loop built to keep you using it. eignex.com/posts/writing-the-loss-function/ #AI #LLMs

"Next-token predictor" is the wrong (or at least incomplete) mental model for post-trained LLMs. I hadn't thought of it like that before. It's always good to understand the tools you're using. https://gmcgoldr.github.io/2026/09/04/llm-next-token-predictors.html

Claude Fable 5.1 is good so far, but I am still avoiding the long, complex tasks it is meant for. Running out of credits halfway through a subagent run is still not recoverable.

Is Claude's text watermark just setting the probability of "byte-identical" to 100%?

Very cool paper: arxiv.org/abs/2608.26263 Instead of having the agent's message history appending forever, you have a mutating state that only sees the latest observation. Seems promising for large code bases.

Took a break from Opus to spend some time with Codex. I now get why people are saying Opus is useless. The model is insanely verbose, no matter if you tell it to stop yapping. I will use both for now since I've yet to get Codex to complete long /goals without pausing.

Alibaba's new chip runs Qwen-3.8 27B at 30 tokens/sec, which is quite usable. Maybe Qwen will reach previous DeepSeek levels of cheap? https://wccftech.com/alibabas-tsmc-built-5nm-risc-v-chip-xuantie-c950-now-runs-qwen-3-8-27b-model-natively-unlocking-massive-vertical-integration-tailwinds/

Bashing on GitHub is very fashionable but is there any other platform as generous with CI compute for public repos? No? If there was they would also be DDoSed by vibe coders

I am totally baffled why Claude Code dropped the ToDo list tool. It seems to me the best/easiest way to have the agent remember specs. You can enable it again by `CLAUDE_CODE_ENABLE_TODO_TOOLS=1 claude`... but still why drop it?

ML got this far on open collaboration. Now it's U.S. export controls, Anthropic lockouts, and Alibaba distillation fights. Sad state of affairs for global ML. https://www.reuters.com/world/china/alibaba-ban-claude-code-workplace-over-alleged-backdoor-risks-source-says-2026-07-03/

Anthropic: the worlds most dangerous model too powerful to release. Also Anthropic: we have no idea why we are forced to pull Fable. It's just a misunderstanding.

The best part about new Anthropic model releases is the accompanying weekly limit resets from all the bug fixes.

Cool article about optimizing images in RAG pipelines. Precomputing image-to-text is obvious, but they have some nice extensions that are worth reading about. https://www.kapa.ai/blog/how-we-index-images-for-rag

Spent the afternoon watching Claude confidently diagnose a Kotlin/Wasm compiler bug: minimal repro, version bisect, "we should pin the toolchain and file upstream." It was a latent infinite loop in my own solver. It's never the compiler. It's never the compiler. Scarily human-like tbh. 🙃

Good HN blog article on using LLMs at a slower pace: using multiple agents on the same PR to surface issues, then reviewing and filtering the comments yourself, is a pretty sensible way to trade speed for better code. https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/

I'm still thinking about the result by OpenAI solving a real conjecture. How much of it is LLMs capability of being good at everything vs doing something really inspired? Or that they probably tried pretty much all major unsolved problems and got lucky on one? Or maybe I'm coping 🙃

Google reportedly auto-updated Antigravity from an IDE into a standalone chat app, broke side-by-side installs, and had people purging binaries to get the old IDE back. Good writeup here. https://www.0xsid.com/blog/antigravity-bait-n-switch

OpenAI must be pissed they just got passed by Anthropic in valuation right as they're gearing up for an IPO. Great timing. https://www.cnbc.com/2026/05/20/openai-ipo-filing.html

Is this the new exciting world of SEO, hacking Google's AI Overview so one false claim gets repeated at scale? Pairs well with Google making AI Overviews the default. https://www.bbc.com/future/article/20260519-google-tackles-attempts-to-hack-its-ai-results

Every time Google's pronounced dead in the AI race, they release something like Gemini 3.5 Flash today. It beats some flagship models while still being the fast, cheaper tier. Still odd they don't match OpenAI or Anthropic on release cadence given their cash flow.

ByteDance just released a 3B unified model that's competitive across image, video, editing, and understanding. Multimodal seems to come for free now? https://huggingface.co/bytedance-research/Lance

Anthropic acquired Stainless, the SDK and codegen behind its client libraries. Points in a familiar direction: more of the API layer pulled in-house, less neutrality in the tooling stack, tighter control over how developers integrate. https://www.anthropic.com/news/anthropic-acquires-stainless

Claude 500 errors again. Maybe I should spend the time to properly investigate OpenCode. So I can at least switch provider when one is down. If only I could run claude on it 🙄

SWE profession is saved, I found something that I'm better at than Claude. It tends to say 1 day of work and finish in 5 minutes. My estimates are typicall better but wrong in the other direction.

Progress bar on compacting.. amazing progress. But seriously, why don't we have an option for amortized background compaction? Sometimes there's just a stream of related simple things to do and I can't be bothered to micro manage context.

I'm returning to an old codebase after 6 years. I wish I had someone else to blame for all of this. The lack of LLM comments feels refreshing… though they would've been useful.

Released v1 of Kumulant: a Kotlin/KMP library for streaming statistics. It summarizes data streams without buffering, using sketches, histograms, and rolling moments with tunable concurrency. github.com/eignex/kumul... #Kotlin #KotlinMultiplatform #DataEngineering #StreamProcessing

Conventional wisdom: typed schemas and serialization formats are separate concerns. I tried. They wanted to be the same thing. Designing one shape for both typed Kotlin and YAML deleted a translation layer I'd quietly planned to build. eignex.com/posts/one-sh...

Hi, I'm Rasmus. Head of engineering by trade, building Eignex on the side: open-source infrastructure for optimization loops in production. Posting here as I build it. eignex.com/posts/building-eignex-in-the-open/ #buildinpublic #opensource