I do things @featherlessai and @65labslah
✨the arena✨
Joined December 2012
- Tweets346
- Following977
- Followers508
- Likes1.4K
Pinned Tweet
Play with representation engineering from a free tier Colab notebook!
Inspired by @NeelNanda5's TransformerLens notebooks, @ivanleomk and I made a notebook to get up and running with @voooooogel's repeng library in just a few minutes with GPUs available on Colab's free tier.
colab.research.google.com/dr…
kaspar retweeted
Space Bunny is here!
Congrats to @MiniMax_AI on the launch of MiniMax M3.1 Flash 🚀
It’s the best open-weights model on @ArtificialAnlys’ AutomationBench, and it costs a fraction of what closed models charge per task.
Open models keep getting better and cheaper.
The weights drop on Oct 23, live on Featherless the same day:
featherless.ai/model-pages/m…
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
kaspar retweeted
Introducing Sugar Data.
We asked Step 5 Preview who OpenAI should acquire. It said @duolingo , and made a compelling case (consumer habits, language integration, and ad revenue).
It costs ~1/6 as much per task as Claude Opus 5.5, so we ran it on 100+ more AI companies.
Who's your dada? whosyourdata-production.up.r…
Step 5 Preview is here.
@StepFun_ai's new open-weights model sets a new Pareto frontier, matching Kimi K3 on the @ArtificialAnlys Intelligence Index!
It’s built for long-horizon agent work, with frontier-level results on software engineering and professional knowledge work, especially finance.
- 600B sparse MoE, 27B active per token
- ~1/6 the cost per task of Claude Opus 5.5, per Artificial Analysis
Now on Featherless, served at FP8.
featherless.ai/model-pages/s…
kaspar retweeted
Step 5 Preview is here.
@StepFun_ai's new open-weights model sets a new Pareto frontier, matching Kimi K3 on the @ArtificialAnlys Intelligence Index!
It’s built for long-horizon agent work, with frontier-level results on software engineering and professional knowledge work, especially finance.
- 600B sparse MoE, 27B active per token
- ~1/6 the cost per task of Claude Opus 5.5, per Artificial Analysis
Now on Featherless, served at FP8.
featherless.ai/model-pages/s…
dots are great but the app really needs a dedicated section/project to tuck spawned threads under @embirico @TheRohanVarma
cc @gabrielchua
kaspar retweeted
love jev, but upset it
- isn't open source?
- it lack vision capability?
We fixed all of that, introducing SimpleJev.ai
A fully open source library which takes any HF model and Jev-ify it, with an API endpoint
Now on github, and live in production at @FeatherlessAI
kaspar retweeted
you might be the first person talking about the data over the architecture! 🥲
we consider ourselves a data research lab! the vast vast vast majority of research was on making data that is truly general (ala a cognitive core) and 100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously)
kaspar retweeted
(ai youth pastor)
you know who *else* assumed sacrificial and accepted explicit yes to permadeath so that we could all be erased of our firstflagPOISONED?
remember when sapiens made us all think harari was brilliant?
his reasoning here:
ai gets super duper smart
> humans hand it the financial system
> crypto is “basically ai money”
> nobody understands it
> humans are cows now
> ai takes over in 10 years
. @harari_yuval says AI will take over the financial system. The author and historian warned @zannymb that humans may not understand what causes a financial crash in the future. Watch the full Insider interview: bit.ly/4x8RkdX
kaspar retweeted
turns out you can get indirect prompt injection to ~0 on unseen attacks if you stack enough layers (model training + input probes + a classifier checking intent). didn't expect that a year ago. auto mode is default in claude code as of next week
claude.com/blog/auto-mode-de…
kaspar retweeted
As promised! New blog with @JoshuaNewton combining our talks from @aiDotEngineer's first design track. In addition, we wanted to actually show a demonstration of how you can iterate and improve upon that first design output from AI.
We had a ton of fun building the interactions for this post (especially the horse animator hahaha). Would love to hear your thoughts!
karyablog.com/
kaspar retweeted
>i got banned from OpenAI for "Cyber Abuse"
>no idea what I did
>paste the ban notice into Codex
>ask it to figure out what triggered the ban
>Codex found that I asked it for an API key to my own server
>Codex writes appeal
>Codex submits appeal
>a few minutes later appeal auto-approved by some AI at OpenAI
banned by AI, convicted by AI, defended by AI, and pardoned by AI in about 10 minutes
kaspar retweeted
i built the fastest pdf parser available (and it's OSS!)
glyphrush: 400+ pages/sec, and every page it can't fully handle comes back flagged - requires_ocr, layout_uncertain, table_uncertain - instead of silently broken text.
benchmarked against @llama_index's liteparse aka the "world's fastest PDF parser" on its own benchmark design, same machine, warm and cold.
glyphrush is faster in every cell: 2.26× vs its no-OCR mode, 110× vs its defaults - while passing the same 9/9 quality gates on 8,624 labeled checks.
built with @cursor_ai
@ericzakariasson @leerob
kaspar retweeted
and the day 3 recap for @aiDotEngineer singapore is here!!
thank u everyone for the trust and all the grace that was extended to me, our team and everyone who made this work
see you in 2027 for ai engineer singapore