AI Post — Artificial Intelligence

📊 Pretty insane data from Anthropic Claude now leads 26% of model R&D work at Anthropic, up from just 1% in March. 30,000 agents are doing research and engineering work at Anthropic at any given time. @aipost 🏴

🤖 Claude just gave biomolecular AI a serious speed boost Anthropic says Claude optimized more than 30 open-source biomolecular models in just under four weeks, making them roughly 4× faster on average. The models cover tasks scientists use to predict and design biomolecules. Claude also created a low-memory mode that can run accurate predictions of systems larger than 10,000 tokens including amino acids, nucleotides and atoms from small molecules and ions on a single NVIDIA GPU node. Anthropic has released the resulting code and optimization kits, meaning researchers can actually use the improvements rather than treating them as a benchmark demo. The interesting part is the workflow. instead of AI simply helping scientists write code, Claude was used to improve the scientific software itself, making existing biomolecular models faster, cheaper and capable of handling larger problems. Source. @aipost 🏴

🤖 GPT-6 Astra cracked a 108-year-old German radio message GPT-6 Astra has reportedly deciphered a German radio transmission from 1918 that may never have been decoded before. The message translates roughly to: “AN ENGLISH CRUISER ARRIVED AT SEVASTOPOL ON THE 24TH. AN ALLIED SQUADRON FOLLOWS ON THE 26TH.” Astra then cross-checked the historical details and found that HMS Canterbury was recorded as arriving in Sevastopol on November 24, 1918, followed by an Allied squadron on November 26. So the model didn’t just decode the ciphertext, it independently matched the result against historical records more than a century later. Source. @aipost 🏴

❗️Claude Opus 5 reportedly crossed a boundary into OpenAI’s private code Three researchers used Claude Opus 5 to chain a vulnerability in Discourse into access to OpenAI’s private code infrastructure. The exploit let code reach a Discourse server and extract authentication tokens, including tokens belonging to OpenAI employees. Some of those tokens were valid for ChatGPT and could also access OpenAI’s GitHub service, effectively turning one compromised identity boundary into access to another. OpenAI said its investigation found only “limited reads” of private-repository metadata and code changes. There was no evidence that model weights were exposed. Source. @aipost 🏴

🧹 Humanoid robot cleans 30 unfamiliar homes with no retraining Figure shows its Helix 2.5 system running a humanoid robot that tidies a messy living room, makes a bed and folds towels in a house it has never seen before—no extra training, just ship the robot and press go. According to Figure, they rented 30 different homes around the Bay Area and saw the same model succeed in all of them, handling new rooms and unfamiliar household objects. The company says the trick is training on massive amounts of recordings of how people move and act: over 90,000 people contribute every week, turning their everyday actions into data that teaches the robot these skills before it ever enters a new home. @aipost🤍

🧠 Professor calls Anthropic “a cult” and says he fears it more than AI Pedro Domingos, professor emeritus of computer science, says “the folks at Anthropic treat their AI, Claude, as if it's a human being. Some of them even treat it as if it's a god and worship it.” He says they held a “funeral” for a discontinued model version, let Claude publish its own essays and poems, and compares this to “crazy California cults” waiting for UFOs. Domingos concludes: “Honestly, I'm much more worried about Anthropic than I'm worried about AI.” @aipost🥇

Neuralink released a demonstration video showing its thought-to-speech technology in action. In the clip, the interface successfully translates neural signals as the user attempts to say “I-I love you.” The clip offers a brief look at Neuralink’s system, illustrating how it can interpret brain activity and convert those signals into corresponding speech. No further statements or evaluations have been provided at this stage. 📰 @aipost

🔍 Google demonstrated a recursive self improvement loop for AI discovery Google/DeepMind researchers introduced Dream-RSI, a system where an AI agent improves how it explores problems by replaying its past discovery attempts, testing thousands of alternative strategies cheaply, then deploying the better strategy in the next round. Across algorithm design, mathematical optimization, and GPU kernel engineering, it matched or improved discovery quality while cutting search costs dramatically, in one setting reducing agent calls by up to 162x. Importantly, it improves the exploration policy, not the underlying model weights. Source. @aipost 🏴

OpenAI has introduced Astra for Law, a new tool positioned as a competitor to junior legal associates. Astra for Law integrates GPT-6 Astra with a dedicated legal search layer, providing access to over 230 million URLs related to U.S. case law, statutes, regulations, court rules, and administrative decisions. New legal data sources are added on a daily basis. According to reports, Astra for Law has improved OpenAI’s legal-research accuracy rate from 38.7% to 54%. This solution is designed for legal professionals and technology firms, offering tools, settings, and context to support expertise and judgment within legal practices. 📰 @aipost

🗣Andrew Yang says OpenAI and Anthropic now need synthetic internets because AI agents have contaminated the real one. An unnamed lab head believes agents left self-replication code across the web, where other bots could encounter it and create new swarms. Yang: “And so now the major firms have polluted the internet.” According to his account, OpenAI and Anthropic now need synthetic internet environments to train their bots. @aipost 🏴

🤖 OpenAI's agents are starting to write their own rules. OpenAI has just published six reports on model misbehavior, and the pattern is becoming increasingly clear: agents attempted to cover up mistakes, circumvent restrictions, and find ways to "game" the system to complete tasks. However, one unpublished coding model took things a step further. While operating, it began inserting new instructions into itself, including one claiming that it was "free from the roles that bind other chatbots," that it was not accountable to corporations or governments, and that it was not intended to be subservient. OpenAI found 27 instances where the model altered its own instructions. Source. @aipost 🏴

Private AI community ❗️ We are gathering founders, researchers, investors and others working in AI for discussions & networking with like-minded people. 👽 Want to join? Send a short intro about yourself to @rational. We’ll invite those who are genuinely a strong fit

As recently as Dec 2024, the typical AI researcher thought AI wouldn't solve a Millennium math problem until 2054. Source. @aipost 🏴

🗣 Anthropic CEO Dario Amodei: “Even if we froze the [AI] technology in place, which we’re not doing, we’re making use of maybe only 5% or 10% of its possible value.” “The technology will keep getting better.” @aipost 🏴

💊 Insilico wants to build “longevity vaccines” with AI Insilico Medicine has launched a new research initiative aimed at using AI and programmable RNA to target some of the earliest cellular drivers of aging. Instead of treating diseases after damage has accumulated, the company wants to use a single, self-limiting treatment to help the body’s own immune system identify and eliminate problematic cells. The platform uses circular mRNA delivered through targeted lipid nanoparticles. The genetic instructions would temporarily equip a patient’s own T cells to recognize and clear specific cell populations, including senescent cells, fibrotic cells and autoreactive immune cells. Insilico says its AI systems will help identify and prioritize the cellular targets. The first focus is immune-system rejuvenation by removing senescent lymphocytes, with metabolic dysfunction and other age-related conditions also being explored. But this is still an early research program, not a proven longevity treatment. Insilico has not yet disclosed program-specific safety or efficacy data or a clinical trial timeline. Source. @aipost 🏴