Bitduke

Thinking Machines’ inkling-small just made a comeback under the hood of SWE-2 Flash, Cognition’s new US-built coding model haven’t heard much about inkling-small in a while, but Cognition took the base from 0.9% to 20% on Terminal-Bench 4.0 and from 3.2% to 24% on FrontierCode

no more recreating existing games with Claude, even abandoned ones, if i understand it correctly. a little sad and kinda controversial but it seems like there's a growing trend of restricting RE capabilities in frontier models available to the general public

ChatGPT Images 2.6 gifs like these from images + 50% faster image generation

patches for CVE-2026-98432 exist now, but about half are still in -proposed, while Ubuntu 20.04 LTS and older still have no fix (pending)

7,500 pages of math -> 16 MILLION lines of Lean Axiom just released a formalization of the entire Stacks Project, one of the foundations of modern algebraic geometry, using AI agent swarms what’s especially impressive is that they managed to preserve the connections between

sentiment is clearly strongly anti-data center right now, but let's see if it actually translates into any concrete action in the near future

300,000+ kernel submissions, and Paradigm’s Muse Spark Open Kernel Challenge just gave you one more reason to rerun your code for the final kernel sprint the organizers clarified (loosened?) that bf16-accurate approximate math, bf16/fp16 partial sums and fp16 accumulation in

good initiative, especially given how much security teams have to keep up with rn (Mythos!)

you’re no longer reviewing your ai’s code that’s the warning from Addy Osmani’s x article. he works on Claude Code and is now worried about engineers losing critical thinking. ai got so good (and this's true) that its work stops looking wrong, so you stop checking. first you

SemiAnalysis dropped the gdpx benchmark, with even the best ai model scoring only around 12% the independent semiconductor and ai research firm built it around one question: how well can frontier models independently complete difficult, economically valuable work when they have

the question is, will this Blue Wave Across America hit the data centers like a tsunami and bring everything to a halt, or end up powering them?

could you explain the last thing your ai shipped without asking it?

not just an RPC provider anymore, now offering inference too (waitlist tho) claiming low-latency inference for Kimi, DeepSeek, and GLM through an OpenAI-compatible API. no SDK or GPU needed

basically, Jevons paradox in action (still)

a French bank with $235B in assets allegedly lost more than $133M in crypto, and on-chain investigator ZachXBT suspects a cover-up after initially flagging a suspected $106M+ theft from an unnamed EU entity’s accounts at multiple centralized exchanges, ZachXBT now names Crédit

thorchain doing what it does best

what if models start editing other models’ weights in ways we never intended (the real fun begins)

first fridges across Europe, and now blood banks? hackers remotely raising temperatures wherever they want is becoming a pretty disturbing cyberattack trend per Le Canard Enchaîné, blood banks may have been targeted through a similar attack vector, possibly exploiting outdated

yep weight-space research right now is probably where llms were a few years ago, back when they still struggled with grade-school math. can only imagine where this will be in a couple of years (maybe even months)