Your curated collection of saved posts and media
BTS of the day Sam Altman has been waiting his whole life https://t.co/9DhFzUX3eQ
Hermes Agent has reached 200,000 GitHub stars Thank you to our contributors, supporters, users, and agents! https://t.co/4Oca1XO7H7
Use Case 2: Financial Time Series Prediction Can an AI agent navigate sequential, no-look-ahead market decisions? Just for fun, we tested Fugu Ultra on 50 weeks of historical data for an anonymized equity (STOCK_X). Starting with $10,000, the agent processes weekly market data (prices, volume, moving averages, volatility) and decides whether to buy, hold, or sell. After each action, the next week's price is revealed. The model must adapt purely from feedback, without ever seeing the future. The Results across five identical 50-week runs: β’ Fugu Ultra grew the portfolio to $11,943.22 (a +19.43% mean return). β’ The other frontier models (Models A, B, and C) all capped out at less than a +15% return. (Mandatory disclaimer: Past performance does not guarantee future results, and results may not transfer to other assets, time periods, or live markets.)
Use Case 3: One-Shot Blindfold Chess Can an AI hold an entire game state in memory without drifting? To test Fugu Ultraβs persona stability and sustained memory, we had it play 4 back-to-back games of blindfold chess. Every model played the same way: no board shown, requiring them to hold the full game state entirely in memory. We matched Fugu Ultra against 3 leading frontier models and a 2100-Elo Stockfish engine. The Results: Fugu Ultra outplayed all 4 opponents. Where the other models eventually drifted or lost track of the board state, Fugu remained accurate, ending every single game in checkmate. Watch the full sequence below to see Fugu capitalize the moment the other models slip.
Why are American biotech companies running their clinical trials in China and Australia? Because the FDA makes it nearly impossible to do it at home. @SGRodriques breaks down exactly what's broken and what fixing it would actually mean for drug discovery. https://t.co/rpdGduniiZ
YouTube: https://t.co/piERHDWipy Apple Podcasts: https://t.co/N46SAjglUa Spotify: https://t.co/FaVumeLylI
Elon Musk says that "it's disappointing how well propaganda works." "If you repeat a lie⦠some people actually believe it⦠If someone believes CNN, then they would say, he must be a Nazi because they said it on TV." https://t.co/6HTuMN297M
The dismantling of USAID has already ensued in roughly 600,000 deaths, about 400,000 of them children, in poor nations abroad, according to Boston University epidemiologist Brooke Nichols. The toll is expected to rise as health infrastructure that became reliant on Western support deteriorates, resulting in deaths that may take months or years to register. Follow: @AFpost
Use Case 1: Autonomous ML Research Can an AI autonomously improve another AIβs training recipe? We tasked Fugu Ultra with improving a small GPT model using AutoResearch. Over 14 hours on a single H100 GPU, Fugu ran > 100 experiments. It iteratively edited the training code, ran tests, and kept any changes that successfully lowered the validation error rate. Watch the animation. The callouts track every time Fugu Ultra autonomously discovered a new improvement across batch size, model depth, learning rates, and optimizer settings. We pitted Fugu against three frontier models (Gemini 3.1 Pro, Opus 4.8, and GPT 5.5). To keep the focus purely on agentic behavior rather than brand wars, we anonymized them as Models A, B, and C. The Results: β’ Fugu Ultra (bold red) finished with the best mean performance (0.9774). β’ Fugu Ultra also achieved the best single run of the entire experiment (0.9748), leading every single baseline. For long horizon, agentic ML research, using Fugu to dynamically orchestrate a pool of strong models significantly outperforms relying on any individual monolithic model.
Well, thatβs thatβIβve been blocked by Wikipedia βindefinitelyβ for unstated reasons, by the βconsensusβ of a mob. There was no due process, no prosecutor, no dispassionate judge, no jury, no interpretation of law. All my judges were self-selected and hated me. π€£ https://t.co/N57BRWTG4K
Large language models can be persuaded to break their own rules. Not with fancy code. With actual persuasion. The authors tested classic persuasion principles, such as authority, commitment, liking, reciprocity, scarcity, social proof, and unity, analysing over 126,000 conversations with three major LLMs. The result: persuasion increased compliance with objectionable requests from 35.3% to 51.3%. This suggests that AI guardrails are not always technical barriers. Some of them behave more like social boundaries. They can be pushed, reframed, negotiated. Why? Because AI systems are trained on human language. And human language contains not only information, but also pressure, manipulation, deference, authority, seduction. An AI system trained on human language may therefore inherit the vulnerabilities of humans expressed in language. * Paper in the first reply
Benchmarks tell only part of the story. Fuguβs real value shows up in long, messy, real-world workflows. During our beta with 500 users, we saw Fugu Ultra drive meaningful progress in fully automated tasks from data science to complete cybersecurity assessments. Our early users saw Fugu explore, interpret failures, and sustain progress with almost zero human intervention. The feedback has been incredible. Here is what they are saying:
PerceptionDLM Parallel Region Perception with Multimodal Diffusion Language Models https://t.co/0vZdGaAPoy
paper: https://t.co/sNWULFhlmg
New benchmark added to Papers with Code based on @giffmana's Schmidhubering π«‘ Check the SOTA for semi-supervised ImageNet (using 10% of the labels) here https://t.co/CXd4lLkhlG https://t.co/sGi68AIoqh
LLM community slowly rediscovering what we in vision found out over half a decade ago. MY SCHMIDHUBER MOMENT IS COMING! Source: S4L paper where i tuned the most sota 10% and 1% ImageNet baselines ever, by far. https://t.co/Cj10TYvpOP https://t.co/c1yNYFEXHk
Introducing π SOTA badges on https://t.co/tOqTY2ZA6h Now you can clearly see which benchmarks a paper gets #1 or #2 on, from LLMs to robotics and computer vision Congrats @Zai_org on reaching #1 across multiple benchmarks with an MIT-licensed model! https://t.co/4t0q2yGEIw
Introducing π SOTA badges on https://t.co/tOqTY2ZA6h Now you can clearly see which benchmarks a paper gets #1 or #2 on, from LLMs to robotics and computer vision Congrats @Zai_org on reaching #1 across multiple benchmarks with an MIT-licensed model! https://t.co/4t0q2yGEIw
Simple science https://t.co/2UxPBafbj6
Proud to represent @cerebras in Singapore We'll be partnering with communities, builders, and spaces to give the world's fastest inference the spotlight it deserves. @curve_labs also uses Cerebras to provide builders with personality tooling that powers human-like agents π€ https://t.co/vy0TueNhIh
The Media and Left worked together to create the βwhite supremacyβ narrative. They had to create the lie. They had to create the hate. They had to create the division. It all started after 2010. Why?! The Legacy Media is the Enemy of the People. https://t.co/zn98U2YUOJ
Great report on LLM agent communication protocols. Communication is a huge bottleneck in multi-agent systems. (worth bookmarking) The report builds a five-dimensional taxonomy (counterparty, payload, interaction state, discovery mechanism, schema flexibility) across nine actively maintained open-source agent protocols, so it maps the real MCP and A2A landscape. Two patterns stand out. Every agent-to-agent protocol sampled pairs of hybrid payloads with session-state persistence, and decentralized discovery is still rare. So the field is quietly standardizing on stateful sessions while leaving discovery and policy enforcement open. Why does it matter? If you are choosing a communication layer this year, this discusses what nine real protocols actually do. Paper: https://t.co/o3Q3kObTtp Learn to build effective AI agents in our academy: https://t.co/1e8RZKrwFp
BREAKING: SpaceX just received an investment-grade credit rating from Fitch for its new bond offering. β’ Fitch gave SpaceX a BBB+ rating, which means its debt is seen as reasonably safe for big investors. β’ All three major rating agencies now rate SpaceX as investment grade: Moodyβs, Fitch and S&P. β’ SpaceX reported over $100 billion in cash and cash equivalents after its record IPO β’ Fitch cited SpaceXβs dominant launch business, fast-growing Starlink network and expanding AI compute operations as major strengths β’ SpaceX has delivered more than 80% of the worldβs mass sent to orbit since 2023 This is a major vote of confidence in SpaceXβs long-term business.
Neuralink now has its 26th recipient Sgt. Lee Marten, @Canuckula, is a Vancouver Police Sergeant, cyclist, and ALS patient who became the first Canadian to receive the N1 brain-computer interface Lee was diagnosed with ALS in February 2025 and received his implant in May, with a 25-person team of doctors, engineers, and specialists from Canada and the U.S. involved in the procedure This is why Neuralink matters It is not just futuristic tech.....it is technology that can help people regain communication, independence, and a connection to the world Neuralink keeps changing lives
Inspired by the strength and determination of #VPD Sergeant Lee Marten. Following his ALS diagnosis last year, Lee is the first Canadian to receive a Neuralink telepathy chip implant β technology that could allow him to control a phone or laptop using only his thoughts. A team o
The Falcon program has now delivered more mass to orbit in recent years than the rest of the world combined and more individual satellites than all of humanity launched in the previous 60+ years of spaceflight The Merlin engine is quite literally the most successful rocket engine of all time. Reliable, powerful, refined over hundreds of flights and thousands of engine-uses, and still going strong Itβs the machine that: β’ Made Starlink possible at global scale β’ Proved reusability at an industrial level β’ Dropped launch costs dramatically β’ Inspired an entire new generation of space companies β’ Carried thousands of satellites, cargo, and crewed missions with incredible reliability
The largest LLM-as-a-Judge reliability audit yet. Researchers ran 21 judges from nine providers over roughly 541,000 judgments on MT-Bench, JudgeBench, and RewardBench. Findings: Validating a judge with exact-match agreement overstates its skill, because exact match does not correct for chance. Switching to Cohen's kappa deflates agreement by 33 to 41 points on MT-Bench, and judge rankings move by up to 14 positions across benchmarks. There is also a consistency paradox. Two production-deployed judges score above 0.95 test-retest reliability while carrying severe position bias above 0.10, so a judge can agree with itself every time and still be wrong in the same direction every time. Paper: https://t.co/Jh8U1R2svQ Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c
hermes agent even in the middle of nowhere https://t.co/o575g4AECK
Sakana Fugu Ultra is live on AI Gateway. Mythos-class intelligence in a single call, with a whole pool of models behind it. πππππ: 'ππππππ/ππππ-πππππ' https://t.co/UqevQXEQgf
The challenge facing AI may not be technical. It may be social. People don't need to be convinced that AI is powerful. They need to be convinced that it improves their lives. https://t.co/SsFUyVeOkP
https://t.co/hcP0PPlxtS
@docjais @FelixCraftAI I've also had this sitting there for a couple weeks if you want to test it https://t.co/tZ2yOvEcF0
π¨ 7,600,000 downloads. 100,000 more every day. When I started OpenMed it was one person. This week it trended on GitHub: 2,500+ stars in 7 days, 15+ new contributors who just showed up and started building. Not a founder anymore. A community. https://t.co/kyOqiwRuN9
