Pinned Tweet
This account is now The AI Engine Room.
The goal is simple: bring you the best information on AI and help you stay up to date. Model launches, agents, tools and local AI, explained clearly.
Subscribe to our channel: youtube.com/channel/UCLBIAlp…
2
1,216
METR says frontier agent time horizons have doubled about every 3.6 months. What stands out to me is their finding that agents still judge worse than human experts when feedback is slow. Which gap closes first, speed or judgment?
We've been working on understanding model capabilities beyond what our prior time horizon tasks could capture.
Today we’re sharing a research update. We find that AI R&D-relevant capabilities are improving rapidly and are getting harder to measure.
1
7
Source: METR research update, Nov 20. Doubling time 3.6 months (95% CI 2.1 to 4.9) across their TH1.1 and TH2 tasks. nitter.gq/METR_Evals/status/2123…
We've been working on understanding model capabilities beyond what our prior time horizon tasks could capture.
Today we’re sharing a research update. We find that AI R&D-relevant capabilities are improving rapidly and are getting harder to measure.
2
Google priced Gemini 4 Neon at $3 per million input tokens and $15 per million output, with a 1M token window. The part I care about is the claim that it cuts redundant file reads in long agent sessions, since that is where agent runs get expensive. Has anyone measured it on a real repo yet?
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
3
27
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
6
OpenAI's GPT-6.2 Astra is reportedly the first externally deployed model with a High rating for AI self-improvement. It also missed two of three second-stage research engineering cutoffs.
So the rating says it can help with research, and the test says not unsupervised yet.
Would you trust it with a research task without checking its work?
2
22
OpenAI says its models found up to 14% better power, performance and area on AMD's flagship GPU IP in under five months, on designs AMD engineers had already tuned for years. I think the interesting part is that it was verified, not just proposed.
Would you trust a model to touch chip design that's headed to manufacturing?
Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing.
The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon.
New evaluations also show our models getting better at certain chip design tasks.
We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers.
openai.com/index/advancing-c…
2
17
Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing.
The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon.
New evaluations also show our models getting better at certain chip design tasks.
We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers.
openai.com/index/advancing-c…
3
Tencent just open-sourced Hy4 flash. It's 224B parameters but only 12B active per token, with a 1M token context window. I think that active count is what makes open weights realistic for small teams now.
Would you self-host a model like this, or stay on an API?
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base.
224B (12B active), 1M context, natively multimodal.
Fast, efficient and built for everyday productivity.
It's yours now. And stay tuned for Hy4 coming up next!
🤗: huggingface.co/tencent/Hy4-f…
🐙: github.com/Tencent-Hunyuan/H…
📄: hy.tencent.ai/research/hy4-f…
Released under the Tencent Hy License.
See full terms here: huggingface.co/tencent/Hy4-f…
1
14
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base.
224B (12B active), 1M context, natively multimodal.
Fast, efficient and built for everyday productivity.
It's yours now. And stay tuned for Hy4 coming up next!
🤗: huggingface.co/tencent/Hy4-f…
🐙: github.com/Tencent-Hunyuan/H…
📄: hy.tencent.ai/research/hy4-f…
Released under the Tencent Hy License.
See full terms here: huggingface.co/tencent/Hy4-f…
2
OpenAI launched ChatGPT Wallet, and every agent purchase needs your approval through Face ID or a fingerprint.
It reportedly needs a Pro plan from $100 a month. Approval is covered. What isn't clear yet is who pays when an agent buys the wrong thing.
Would you let an agent check out for you, or keep that call yourself?
1
20
OpenAI says ChatGPT for Healthcare now has more capable computer-use agents and 20+ new healthcare plugins. Cigna and CVS Health are on the list of organizations using it for prior authorizations, referrals and discharge planning. Paperwork is where I'd expect agents to earn their keep first.
If you work in healthcare, would you let an agent handle prior auths?
ChatGPT for Healthcare is advancing with more capable computer-use agents, stronger clinical answers, and 20+ new healthcare plugins.
@AbridgeHQ, @AmbienceAI, @Cigna, @CVSHealth, and other leading healthcare organizations are putting our models to work across prior authorizations, referrals, discharge planning, and patient outreach.
Together, these advances connect medical evidence with practical help across the tools teams already use—so they can spend less time navigating systems and more time caring for patients.
openai.com/index/advancing-c…
1
11
Alibaba shipped Qwen4-Max overnight. Their own numbers say it beats Qwen3.8-Max with up to about 50% fewer tokens and over 60% lower cost per task, at $1.5 per million input and $4.5 output. Those are internal evals, so I'll wait for outside tests.
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
3
1
48
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
3
Weekly AI recap, Nov 9 to 15. Five stories that mattered to me this week, one line take on each. 🧵
1
18
DeepSeek (Nov 14): app features like search, files and voice are now available in its API, plus an open source chat demo. Less flashy than a new model, more useful for builders.
1
1
4














