Big win for @thinkymachines šš. Community wave been waiting for too long for good new small MoE
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/inkā¦
12
Interesting findings
Replying to @askalphaxiv
A few findings from the agentsā experiments:
⢠A base model recovered 86% of an RL-trained modelās improvement on MATH-500 simply by borrowing its prompt cache, without changing its weights.
⢠Models passed memorized facts about fictional people and their cities to other models through training text that mentioned none of those people or places.
⢠Deleting a factās original KV-cache entries and then training a small readout adapter recovered those facts verbatim in 83ā94% of synthetic-list tests from information remaining elsewhere in the cache.
9
BREAKING: Terence Tao and 24 other Fields Medalists just signed a letter telling AI companies they're destroying mathematics
>headlines: AI solves famous math problems
>25 greatest mathematicians respond today
>"we are witnessing a general threat to intellectual work"
>ai labs are treating math problems like a benchmark you can brute force
>āsolutions are announced in a rush, leaving no time for a proper writeup and citing relevant previous work of others"
>they're talking about openai
>"this raises severe attribution and plagiarism questions"
>they're definitely talking about openai
>openai offered to put math professor name on navier-stokes proof
>condition: abandon his coauthor bc he works at anthropic
>"the most precious resources of our profession are students and ideas"
>for the labs the most precious resource is GPUs
A Severe Misalignment of AI in Mathematics
23
Introducing ChatGPT for Dishwashing
GPT-6 achieves state-of-the-art results on dishwashing benchmarks for long-running jobs and difficult stains.
The NavierāStokes Millennium Prize Problem is a $1 million challenge about whether the equations describing fluid motion can break down. In this video, I accidentally found a counter-example while trying to optimize my dishwasher. But in reality, OpenAI used 10,000 AI agents to find a proof in 88 hours: openai.com/index/navier-stokā¦
Parody of the original GPT-6 Astra ad
@OpenAI #ChatGPT_Partner #chatgpt
Credits:
Written and Directed by Joma
Co-Writer - Henry Connor Coan IV @coan_henry
Production by Dauntlus Studios @dauntlus.studios
Producer/Co-Director - Marcus Liew @dauntlus
Camera Operator - Amos Lee @amoselijahlee
Art Director - Alethea Soo @aletheasoo
Art Assistant - Alicia Lim @aliciatkl_
Assistant Editor - GPT-6 Astra
Music: Tensions Run High by Soundridemusic
youtube.com/watch?v=Ly9H63SLā¦
Looks smart AF. Letās close eyes for AI solutions and keep working on what it already solved and mathematicians struggling on it over decades.. letās continue struggling for the hobby it seemsā¦.
Important statement from the Association for Human Mathematics
ahmath.org/statements
13
It does work. The default cut-off message is spliced in raw ā mid-sentence, no space:
`ā¦if main proof is product+1?You are out
of thinking time.`
Crude, but that can be the useful part.
1
41
Qwen3.8's thinking control is a system prompt, not an architecture change. Which might explain why people report better results on medium.
xhigh and low each inject an instruction telling the model how to think.
medium injects nothing. No system message at all.
Full strings in the image.
Qwen3.8-27B ships with reasoning_effort on xhigh. Turns out that's the one setting you don't want.
Ran all four levels through the same config, 1083 tasks each. low and medium tie. xhigh burns 3x the tokens of low, loops 10x more often, and scores no better for it.
I'd just run low effort. Config below.
UD-IQ2_XXS since that's what fits 12GB, but I feel like higher quants will land in the same place.
a metric can look sophisticated and still be structurally blind because of one hidden invariance property
29
How to apply for 1000x reduction on plan cost?
šØ SAM ALTMAN: "A REDUCTION IN COST OF ABOUT A 1000X" š¤Æ
Sam Altman revealed the cost to solve a hard reasoning problem has plummeted by 1,000x since the release of o1 just 16 months ago (comparing o1 to the new 5.4 model).
How is this possible? Altman credits two main factors:
1ļøā£ We're Still in the Early Days: āWe are so new to this paradigm that there is massive room for improvement. We're still doing things in "dumb ways" and getting better fast.ā
2ļøā£ Human Ingenuity is Underrated: āIt's not just the models getting better. It's kernel engineers, power engineers, and data center designers finding brilliant ways to optimize every step of the process.ā
The compounding effect of better models + better human engineering is driving an exponential decrease in the cost of AI reasoning.
44
MC Naive's retweeted
Replying to @trq212
Looks like that bug is still there in 2.1.62. In 8 minutes, how can it use 19% of the Max account?

























