Was at Anthropic HQ this afternoon. A few takes: - Problems in the AI era roughly are two kinds: the ones that are saturating and the ones that compound values. Work on the latter, and build a harness that moves the exponential. - Context anxiety was real at one point: the model refuses to do the work and as they complain about the little left context, not some misconfig or my illusion. and they fixed it in recent models. - Unhobbling should be done with every model upgrade. If the model upgrades and the harness doesn't, the model degrades. - They've got famous ML papers on the wall, plus The Bitter Lesson. Love that. Back at the Jina office, we used to hang our conference posters up. Miss that. - Security around the office is tight.

the charm of local/self-hosted AI is to run abliterated models: Qwen-3.8-27b-abliterated 110~120 tps on ½ nvidia rtx6k (½ bc cheaper and just enough size to run it). image+text, context 230K. possibly the best hill-climbed and actually usable abliterated model on a low budget gpu. the prefilling speed is slower than RadixArk's nvfp4 as their weights and the activations are 4-bit, so every matrix multiply runs natively on Blackwell tensor cores. but this gap shrinks in practice as most of the time cost in my long-horizon tasks is on decoding. if u haven't tried any abliterated models before, you definitely need it. also comparing their reasoning chain side by side vs the regular model, the difference is huge.

As a 911 and Model Y owner, watching the Porsche and Tesla releases back to back today makes me a lot. In the afternoon, after Porsche's Mission S. I thought god Porsche is finally back on track. The deep purring of a combustion engine, the center screen and infotainment gone entirely, quad exhausts. Their PR just says that if you want entertainment, bring a phone mount. One can see Porsche practicing first principles with real engineering chops, in a Germany and a Europe that have lost the plot. In the evening, Tesla's Roadster. Gripping the track at 180 degrees, then lifting off and landing in place. That's when I realized that Tesla isn't even in the same race anymore. Once again, American accelerationism shocked me deeply. Porsche's press release pays homage to the 356, a vintage classic. Tesla's pays homage to two fighter jets. “It’s just a car bro” “we're not a car.” Or are we?

This is insane. I was trying to figure out how it stays glued to the track at that angle, and then it throws this curveball. Just as the German automakers are finally getting back on track with Mission S, Tesla's already left the track and taken off.

"Porsche has deliberately chosen not to include a permanently integrated center display or a fixed passenger display in the dashboard." that's the spirit! all-in physical buttons, no infotainment bs!

on jy-crpg-bench gemini-4-argon-high is better than 3.8-flash but still behind opus-5.5 🥲 it's surprisingly stable across multiple runs tho, which is a good thing.

After Singapore, Miami and Suzhou, we're hosting the 4th EMNLP Birds of a Feather on embeddings, rerankers and small LMs for better search, this time in Budapest! We're looking for 10-minute talks on related topics, whether it's published work, a poster, or work in progress. Jev-related work especially welcome! Registration link below

it's true, M3 ultra with hardware ray tracing can do amazing things, sharing some of my 992.1 gt3rs rendering with ray tracing using metal/mlx. always want to build a better renderer than porsche official one, i think im getting close.

i like this release, especially since it's post-trained on qwen. but this one feels like a bit of a flex. I'd bet the Xbox research team has been doing this with svm/xgboost/liblinear since day one...

took me three trips to SF to finally finish Diogo and swyx's episode on @latentspacepod while driving. Diogo was super, super energetic, almost too hyped, and even got a little emotional at some point. Congrats on the launch again!

ICLR 2027 submission: we benchmarked all frontier models and our model performs... Anthropic & OpenAI today: no you don't. ICLR deadline is 9/25, so i guess no excuse not including opus5.5 and gpt6-luna & sol in the paper.

wow, Opus 5.5 looks very promising! On my CRPG benchmark, Opus 5.5 takes only 40 steps to get out of the spawn house and onto the world map, much faster than Fable5.1. Let's see if it can speedrun past gpt6-astra. 4-hour gameplay livestream starting now:

Model-specific inference engines are basically what everyone autoresearches and hill-climbs the hell out of for TPS after each Qwen release. In the early days I worked with MLX too, but these days I just trust oMLX to do the job.

yeah this actually looks really bad for zhipu/glm given the recent situation, so far no response from their official account neither Zibo @ZixuanLi_ or @louszbd respond, probably their first crisis PR.

thoughts on Typesafe/Jev: I'm surprised that a general-purpose classifier (or discriminative model) can be just as interesting to the public as a general-purpose generative model. I've worked on search and discriminative models for years, so maybe I should have called this upfront, tbf I never thought about making them general-purpose beyond search. Typesafe also positions Jev really well as "System 1", a complement to the generative models we already have. Humble, unassuming, and sidestepping the frontier-lab warzone. Jev's success could be disruptive to today's agentic systems in many ways, such as tool-calling and routing patterns. Over the last three years the consensus was to use a small generative model for routing, tool calling, MCP etc. Time to pause and rethink such architecture, because we'll probably see tool calling and routing move back to discriminative models. Whoever ships the next Jev-like open-weight base model will likely win the community.