Bringing you the best information on AI so you can stay updated. Clear, practical breakdowns of models, agents and tools. Stay informed. Stay ahead.
United States
Joined November 2022
- Tweets2.8K
- Following479
- Followers665
- Likes923
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…
Anthropic just shared FlashProve, research that lets developers verify whether a response came from a particular AI model.
That could help teams check whether the model they tested is the same one serving users. The useful next question is how much time and computing power those proofs need in a real deployment.
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
OpenAI is more than quadrupling its bug bounty for critical findings, with exceptional ones now worth up to $172,000.
It's also bringing outside researchers from Cathedral into its next internal security sprint to attack its defenses with OpenAI's own models. Last sprint, it put 25% of its product engineers on finding and fixing holes with Codex.
Would that payout get you hunting?
Should every lab run sprints like this?
As AI becomes more capable at finding vulnerabilities, defenders need to put it to work before attackers do.
We’re planning our next internal security sprint at OpenAI and partnering with Cathedral Systems, whose researchers will use our models alongside our teams to challenge our defenses. They’ll bring differentiated expertise and approaches to our ongoing work to find and fix weaknesses with AI.
We’re also raising bug bounty rewards, including more than quadrupling rewards for critical findings. Exceptional findings will be eligible for up to $172,000, including bonuses.
We’ll keep advancing our models, using them to strengthen our own systems, and expanding access for trusted defenders.
openai.com/index/preparing-o…
Source: OpenAI's post on the sprint and bug bounty changes openai.com/index/preparing-o…
Meituan, the food delivery giant, just shipped LongCat-2.5, a 1.6 trillion parameter model with about 48B active per token.
What surprised me is the training setup. They built it on Chinese domestic AI chips, and in their own table it scores 38.5 on Terminal Bench v4.0, just above GPT-5.6 Sol at 37.3. Claude Opus 5 still leads at 53.9.
It's live on their API now.
The biggest AI launches go up here every day.
Would you try a coding agent from Meituan?
How much weight do you give benchmarks a lab runs on itself?
Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such as creative design and 3D tasks.
Learn more about LongCat-2.5: longcat.ai/blog/longcat-2.5/
Source: Meituan LongCat blog with the full benchmark table and training details longcat.ai/blog/longcat-2.5/
The best AI model scores 12% on SemiAnalysis's new GDPx benchmark.
It covers 11 fields like chip design, cyber, biology, finance and law, with agents getting up to 8 hours per task inside realistic work setups.
Opus 5.5 leads overall. On finance and legal, basically every model scored 0.
Which field do you think cracks first?
Would 12% change how you hand work to agents?
Source: SemiAnalysis write-up newsletter.semianalysis.com/… and the live leaderboard gdpx.semianalysis.com/
Unsloth says its 1-bit build of DeepSeek V4.1 Flash works on 128GB machines like a MacBook or a DGX Spark.
The download is 233GB, but 110GB of that is a lookup table that can stay on your SSD and load as needed.
In their own tests, the 2-bit build scored within 0.2 points of the 8-bit reference across six benchmarks. The 1-bit build gave up about 4 points.
Practical local AI news goes up here every day.
What would you run on it first?
Would 1-bit be good enough for your coding work?
We’re releasing new DeepSeek-V4.1-Flash-1017 GGUFs, ranking highest across benchmarks.
Dynamic V3.1 leads across nearly all benchmarks: GPQA Diamond, AIME25, LiveCodeBench v6, MMLU-Pro, Aider & more.
Run on 64GB+ RAM.
GGUF: huggingface.co/unsloth/DeepS…
Blog: unsloth.ai/docs/models/deeps…
Source: Unsloth's run-locally guide with the full benchmark table unsloth.ai/docs/models/deeps…
Ten AI launches from this week in one list.
Google had the busiest week with Genie 4, Gemini 4 Carbon, document editing in the Gemini app and PlanAssist. Mistral and NVIDIA both put out open weights.
Which one would you try first?
Sources:
1. mirendil.com/news/introducin…
2. openai.com/index/introducing…
3. deepmind.google/blog/genie-4…
4. blog.google/innovation-and-a…
5. huggingface.co/blog/nvidia/n…
6. nitter.gq/GeminiApp/status/21158…
7. mistral.ai/news/mistral-larg…
8. nitter.gq/humansand/status/21161…
9. github.com/vercel-labs/agent…
10. deepmind.google/blog/speedin…
A radio telescope in Northern California now takes its orders from AI agents.
At the Allen Telescope Array, SETI Institute and Breakthrough Listen researchers type one plain English request. The agents steer the antennas, calibrate the instrument, catch and fix data errors, then plot the result.
NVIDIA's Holoscan stack pulls up to 2.4 terabits per second straight into GPU memory, so the data gets processed as it arrives.
It has already logged the first real time AI detection of a pulsar, in the Crab Nebula.
Would you let an agent run a science instrument overnight?
Which lab gear should get this next?
AI agents are now steering the Allen Telescope Array in Northern California to search for possible alien signals. 👽
Happy Halloween. Sleep tight. 🎃
See how the @SETIInstitute and Breakthrough Listen are using NVIDIA Holoscan and CUDA-X libraries to listen to the skies → nvda.ws/4zgYoGv
NVIDIA's write-up on the Allen Telescope Array setup: blogs.nvidia.com/blog/ai-joi…
OpenAI is putting ChatGPT and Codex on one shared allowance for Plus, Pro and Business from December.
Usage gets counted in tokens, so bigger models and long Codex tasks drain it faster. Instant chat doesn't count.
OpenAI says most people won't wait on resets more than now, but hasn't said each plan's token amount.
Do you use both enough to notice?
Which one would run your budget dry first?
Bad news for ChatGPT users: starting in December, Plus, Pro and Business limits will be shared across ChatGPT and Codex.
OpenAI quietly updated its help page. Usage will now be measured in tokens, so bigger models and harder tasks burn through your allowance faster. Only Instant chat stays free of the limit. Hit the cap and you wait, upgrade or buy credits. Heavy Codex users may feel this one.
Y Combinator Labs put 50 AI agents in Minecraft with three hours to build a working civilization.
The first night took 22 of the run's 34 deaths. After that, the agents mostly hid. Their informal leader, Ada, told everyone to idle in lit shelters until they got short tasks.
That's a very human reaction to getting burned.
Would you call that caution or a failure to adapt?
What would you have changed in the setup?
This one looks really useful.
Vercel Labs released agent-desktop, a tool that lets a coding agent use apps on your Mac, Windows or Linux machine.
It reads each app through the operating system's accessibility tree, so the agent acts on named buttons and fields with no pixel matching. It runs in the background without moving your mouse, and several agents can work at once.
It's free and open source under Apache 2.0.
The most useful new AI tools go up here every day.
Which desktop app would you hand to an agent first?
What would you never let it touch?
Repo and docs for agent-desktop: github.com/vercel-labs/agent…
