Alexandr Wang@alexandr_wang08:30thoughtful review of muse wonder and muse wonder full from the lovely claire vo! https://t.co/VneEUPK6f6#511482
Elon Musk@elonmusk07:58Grok is the best conversational voice model in the world https://t.co/hiT7uSote7#511482
AI at Meta@AIatMeta07:46Watch the full convo with @finkd, @alexandr_wang and @collision: https://t.co/AJ9GHXjB2d#511482
AI at Meta@AIatMeta07:46Muse Wonder and Muse Wonder Full deliver SOTA performance on personal and professional agentic tasks. Learn more about the models on our researcher blog: https://t.co/koe89SLQts https://t.co/mEIYEFD40g#511482
Elon Musk@elonmusk07:23Major threshold passed. Almost all cars will be electric in the future. https://t.co/ZGDptzVv6H#511482
Rohan Paul@rohanpaul_ai05:55🧵4/n Your workspace also opens on your phone or any browser through the Remote. It app, while the analysis keeps running on your computer. AIPOCH has published 125 example projects you can open and inspect, from whether a simple algorithm beats AI at Wordle to a city’s 100-year rainfall record. Use cases: https://aipoch.com/open-science/use-cases#511482
Rohan Paul@rohanpaul_ai05:55🧵3/n Connectors let the agent pull from research databases like PubChem (chemical data) and openFDA (drug approvals and labels) through MCP, short for Model Context Protocol, a standard plug that lets AI call outside tools. Heavy jobs can run on your own lab server or cloud machine over SSH (Secure Shell, the standard secure way to log into another computer), and only the final files come back. Blog: https://aipoch.com/blog/aipoch-open-science-research-packages#511482
Rohan Paul@rohanpaul_ai05:55🧵2/n Download that exact package and inspect the full session here: https://aipoch.com/open-science/use-cases For bigger tasks, a “Plan first” option makes the agent write a step-by-step plan, and nothing runs until you approve it. You also set how much freedom it gets, from asking before every action, to auto-approving file edits only, to full access. Plans survive app restarts, and if a step gets blocked, the steps after it do not start until you sort it out.#511482
Rohan Paul@rohanpaul_ai05:55An Apache-2.0 app with 4.8K GitHub stars is trying to make AI-assisted science as traceable as checksummed, versioned code. @aipoch_ai’s Open-Science is a free, open-source desktop app where an AI agent reads papers, pulls data from science databases, writes and runs Python or R code, and saves the charts and reports, all in a single workspace on your computer. Research is usually split across PDFs, spreadsheets, code editors and terminals, so results lose their link to the data and code behind them. Open-Science keeps that link: every result records its input data, code, software setup and run history, and the whole session exports as a single .science file that holds the conversation, the chosen files and the evidence, minus your API keys. A public example shows it in action: in “AI Math Citations: Do They Check Out”, the session took 5 references from each of 20 AI-written math papers and checked whether each points to a real publication. A direct check against publication records matched 80 of the 100 and found none to be fake, and after notes the researcher added, the final count was 88 matches, 7 small slips like an online date versus a print year, and 5 unresolved. Because the .science record keeps those 2 evidence sources apart, anyone opening it can see which verdicts rest on human notes, including all 6 positive verdicts in a smaller test of whether cited papers really back 10 specific claims. It is model-agnostic, so it works with whichever AI model you choose (cloud providers, custom gateways, or a Claude or Codex subscription), and local-first, so projects stay on your machine by default. The agent runs code only after you approve it, and it ships with science skills like AlphaFold2 (predicts a protein’s 3D shape) that you can extend with your own. It is Apache 2.0 licensed and runs on macOS, Windows and Linux. 🧵1/n#511482
Rohan Paul@rohanpaul_ai05:43Good safety work increasingly means good search. White Circle’s released RISE, an automated tool for testing whether AI image generators can be tricked into making pictures their own rules forbid. It searches over attack strategies for image models. An open attacker model turns each strategy into many prompts, and the strategy gets credit for how many of them get through. That keeps the search learning even against models that block almost everything. The code is open, the models it uses are public, and defenders can run it on their own image models today.#511482
NVIDIA@nvidia05:39Get to know the five companies and what they’re building ⬇️ https://t.co/6i9rydRLMM#511482
NVIDIA@nvidia05:37Cheers to the UK’s AI builders. 🇬🇧 Tonight in London, @JensenHuang announced new investments in @boltz_bio, @ForesightWorks, @hivergeai, @PrimaMente, and Riverlane. From new medicines to quantum computing, there’s plenty to build. https://t.co/NG0OhNfonM#511482
Santiago Valdarrama@svpino05:20Many software people are still dismissing AI as a fad, or because of the water consumption, intellectual property, or whatever. And it's not just one or two people; I'm talking about a ton of folks closing their eyes and repeating "stochastic parrot" with no idea what they are talking about. Sad to watch.#511482
Rohan Paul@rohanpaul_ai04:26Anthropic's new acqui-hire deal. Its licensing Opaque Systems’ technology for running AI on confidential data and hiring some of its engineers, including CEO Aaron Fulkerson and CTO Rishabh Poddar. > Standard encryption guards data on disk and across networks, but a model can only read it once it’s decrypted in memory. > Opaque’s platform closes that gap by running AI jobs inside hardware-sealed environments that produce cryptographic evidence of what happened during each run. > That evidence lets a bank or hospital verify which code touched its records, on which machine and under which rules. That technology began as a UC Berkeley RISELab project about 10 years ago and carried Opaque to a $300M valuation after a $24M Series B in February, with $55.5M raised in total. Anthropic was already an Opaque customer and had published a confidential inference design with Pattern Labs in June 2025, while noting that some AI chips still lack full confidential computing support. Neither company has said whether Opaque’s platform will keep serving ServiceNow, Accenture and its other customers under the license-and-hire structure.#511482
Rohan Paul@rohanpaul_ai03:59Training agents to do real office work requires real company records, so Scale, Mercor, Handshake, Turing and others now pay businesses for their email, chat, tickets and code history. @Secludy has released Who Sells Data, a free directory of 159 companies that sell training data, RL environments and evals to AI developers, with every fact linked to its source and the date it was checked. Unknowns are marked rather than guessed, so an empty cell reads as “nobody confirmed this,” which is exactly what a security reviewer wants to know. Eval builders, see how you stack up.#511482
Rohan Paul@rohanpaul_ai03:49Jensen Huang on Bloomberg says he wouldn’t sell GPUs to AI labs who are irresponsible. (Link to the full video in comment. )#511482
Tomasz Tunguz@tunguz03:28Yeah, I've been saying this for a while. Human Genome Project was interesting from a scientific point of view (barely), but it made next to no impact on actual therapeutic applications.#511482
Rohan Paul@rohanpaul_ai03:23Bloomberg: CoreWeave’s unsecured 2032 bonds fell below 85 cents on the dollar, the price that usually signals distress, yields near 15% on concern about AI borrowing binge Those notes are $1.25B of 9.625% senior debt sold in June, so lenders have marked them down sharply within 4 months of issue. The shares fell as much as 17% today, though the stock remains more than 60% above its March 2025 IPO price. The business runs on expensive debt. CoreWeave buys GPUs, builds data centers and rents the capacity to AI developers, and it funds most of that with borrowing. The interest bill is now bigger than the losses. The company lost $626M in the second quarter while net interest expense reached $640M, driven by heavy capital spending. Revenue is growing fast, but the cost of financing the growth is consuming it. It’s spreading beyond CoreWeave. Investors are treating it as a test of debt-funded AI infrastructure in general. Oracle’s bonds weakened, and AI-linked stocks from Nebius to Dell fell as markets questioned how much debt sits beneath the buildout.#511482
#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482#511482