←
Bitduke52,375 posts
Bitduke avatar

Bitduke

@bitcoinduke

Applied AI & research | models, agents & cybersec | still stacking sats

◷ Joined November 2013▧ 3,949 Photos and videos
Bitduke avatar
Bitduke@bitcoinduke·12m

ai agents can now finish almost HALF of OpenAI's 2–4 hour research tasks without human intervention, up from just 23% in march. that's roughly a 2x improvement in only 7 months (!) OpenAI just published a new report on how AI agents are accelerating its own research and development, and what i find especially interesting is how uneven their capabilities still are. on the one hand, agents can successfully monitor training runs 79% of the time, but that drops to just 25% when building the infra that actually runs trained models, where they have to deal with much more complex performance and reliability issues so while ai is already taking over more of the research workflow, it seems some pretty important engineering tasks remain bottlenecks

Post attachment

We're sharing new data on how AI is changing research at OpenAI. We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area. This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems. Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry. openai.com/index/research-ac…

Bitduke avatar
Bitduke@bitcoinduke·38m

we finally got the final number from Anthropic, and it’s $83 per share with 980 million shares in its IPO, bringing the total offering to a massive $81.34 BILLION (!), making it one of the largest IPOs in history $ANT is expected to start trading on @Nasdaq tomorrow, Nov 19

Post attachment

We’ve priced our IPO at $83 per share, with trading expected to begin on Nasdaq tomorrow. anthropic.com/news/anthropic… Legal disclaimer: A registration statement relating to the Class A common stock was declared effective by the Securities and Exchange Commission on November 18, 2026. The offering is being made only by means of a prospectus. Copies of the final prospectus may be obtained, when available, from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at (866) 471-2526, or by email at prospectus-ny@ny.email.gs.com; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014, or by email at prospectus@morganstanley.com

Negligible Capital avatar
⟳ Bitduke retweeted
Negligible Capital@negligible_cap·5h

*ASCHENBRENNER DISCUSSES PLANS FOR NEW ITERATION OF FUND: FT The effort is being called "Situational Awareness 2.0"

Post attachment
Dany avatar
⟳ Bitduke retweeted
Dany@danysvnt·10h

OPENAI ENGINEER DOESN'T WANT HIS OWN FAMILY TO KNOW WHAT'S COMING WITH AI "I don't think that I have very comforting words for them" "I'd rather they just not find out that much and continue to live very happy lives" juan felipe cerón uribe has worked at openai since 2022. his job is to stop chatgpt from helping people make bioweapons earlier in the same interview he described how the industry is racing: "let's see what it is and flip a coin and see if we're all going to die or we're going to create a lot of shareholder value" asked about human extinction, he said it's "turning out to be really hard to rule out" and he says people at openai talk about these risks less now than when he joined is that love or avoidance?

0:22
clem 🤗 avatar
⟳ Bitduke retweeted
clem 🤗@ClementDelangue·10h

Number 5 trending on the hub! huggingface.co/harness-hub/Q…

Post attachment

Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…

Post attachment
Bitduke avatar
Bitduke@bitcoinduke·8h

interesting take on why reviews of the new Astra are so mixed

never seen opinions so split on a model. so far it looks like the testers who outsource their thinking hate it for being impenetrable while serious engineers and systems people enjoy solving hard problems with it.

antirez avatar
⟳ Bitduke retweeted
antirez@antirez·15h

The hegemony will be of the folks that at the same time embrace AI without removing themselves from the thinking process. Our goal is to move as many people as possible to the upper class.

Bitduke avatar
Bitduke@bitcoinduke·8h

right, don’t take yourself out of the thinking process

He wrote that there is, today, a new "social caste" that is invisible but not less concerning and divisive: who will not accept the tool, or who will accept it *cutting yourself out* of the thinking process, will lose. (cont)

clem 🤗 avatar
⟳ Bitduke retweeted
clem 🤗@ClementDelangue·9h

Yes! Evaluation of models should be a mix of public benchmark, private benchmark, social signals (ex likes on HF) and testing/experiments. Glad the field is finally maturing!

Andrew Curran avatar
⟳ Bitduke retweeted
Andrew Curran@AndrewCurran_·9h

The Trump administration is considering a requirement that US companies publicly disclose any use of Chinese AI models on their websites. They want to discourage the widespread use of models like Qwen without imposing literal bans that would upset the industry and China.

Post attachment
Bitduke avatar
Bitduke@bitcoinduke·9h

no H200 (or Hopper-generation chips) for China? around 600,000 H200 chips are reportedly stuck in Hong Kong, awaiting permission to enter mainland China

Post attachment

Nvidia stops shipping US-approved AI chips to China reut.rs/4eKFpfo

roon avatar
⟳ Bitduke retweeted
roon@tszzl·18h

so epoch and anthropic capabilities index both appear to show 1.7-2x progress after full coding automation in Q2 ish of this year it appears to be a clean trend break rather than a constant acceleration

Post attachment
Bitduke avatar
Bitduke@bitcoinduke·16h

GPT-6.2 Astra can solve math problems nobody had solved before, yet still struggles with relatively simple research tasks pretty interesting results from Epoch AI's independent evaluation, Astra scored 174 on the ECI, tying Fable 5.2, but fell behind on realistic research tasks (65% vs 82%), struggling to identify the core question, iterate and make estimates under uncertainty meanwhile, it set a new record on FrontierMath Erdős, solving 7/68 problems with Lean-verified proofs, including one with no previously known solution. it also managed 12.3 serial reasoning steps within a single output token, compared to just 4.2 for Fable 5.1 (!) so there’s clearly a gap between reasoning ability and actual research performance here, pretty fascinating. maybe part of it comes down to training. like you can reward a correct proof, but how do you reward a model for noticing it’s asking the wrong question? or maybe it’s more about feedback during the task. Lean tells Astra when a proof fails, while a bad experiment can run perfectly and give it a nice result

GPT-6.2 Astra scores 174 on the Epoch Capabilities Index (ECI), on par with Fable 5.2 and joining it above what the pre-Astra trend for reasoning models would have predicted. This is based on testing we did prior to launch using only our own benchmarks. The ECI is continually updated as more benchmark results become available, and so these values may change in the future.

Post attachment
Bitduke avatar
Bitduke@bitcoinduke·17h

recently, i've started seeing more and more real-world benchmarks (like this one) that test frontier models on actual engineering tasks in physical labs honestly, i'm not surprised. as automated r&d keeps getting better, we're moving closer to the next phase, where ai models will be able to conduct physical experiments on their own with minimal human supervision

Welcome to the real world. Today we’re introducing FrontierEE Bench, the first benchmark to test frontier models on electrical engineering tasks at a physical lab bench. Real-world engineering benchmarks should involve real hardware. At our lab, models control instrumentation, run experiments on physical hardware, and decide what to test next. The best models scored 36.7% across a range of EE domains including mixed-signal circuits, power systems, & electromechanics.

0:05
Bitduke avatar
Bitduke@bitcoinduke·18h

google definitely has every chance to compete in the ai race

Post attachment

Google is about to release another Gemini model. Google competing with Anthropic instead of OpenAI was not on my bingo card for 2026.

Robocurve PBC avatar
⟳ Bitduke retweeted
Robocurve PBC@robocurve·20h

GPT-6.2 Astra is the highest scoring model on RoboDojo-RC Tier 1, a benchmark of real robot tasks. It beats Dream-1, the best robotics model we've tested.

0:10