# AI Research Articoli collegati

Il Centro Notizie HTX fornisce gli articoli più recenti e le analisi più approfondite su "AI Research", coprendo tendenze di mercato, aggiornamenti sui progetti, sviluppi tecnologici e politiche normative nel settore crypto.

Weekly Token Unlock: H Token Unlock Reaches 7.9% of Circulating Supply

Weekly Token Unlock Roundup: Humanity Protocol Leads with 7.9% of Circulating Supply This week's major token unlocks feature several notable projects. Humanity Protocol (H) has the most significant unlock, releasing 270 million tokens valued at approximately $19.88 million, which represents 7.9% of its circulating supply. The project is an anti-sybil blockchain network focused on human identity verification and user data ownership. Other projects with substantial unlocks include SoSoValue (SOSO), an AI-powered crypto investment research platform, unlocking 23.45 million tokens worth about $7.27 million. Collector (COLL), a platform bridging physical collectibles with Web3 digital markets, is unlocking 28.86 million tokens valued at roughly $5.57 million. Each project section includes links to their official Twitter and website, details on the unlock amount and value, a brief description, and a visual token release schedule chart. The unlocks represent planned emissions according to each project's vesting and distribution timetable.

marsbit2 giorni fa 11:40

Weekly Token Unlock: H Token Unlock Reaches 7.9% of Circulating Supply

marsbit2 giorni fa 11:40

Claude Starts Independently Designing New Proteins, Efficiency Surpasses Human Experts by Dozens of Times

Anthropic's Claude AI autonomously designed novel protein binders, achieving a 26.8% overall success rate (354 validated designs out of 1320) across 14 out of 15 target proteins—double to triple the industry average of 10-15%. Guided by a detailed 16k-word prompt but making all design decisions itself, Claude orchestrated existing tools like RFdiffusion and ProteinMPNN over 24-48 hours per target. Notably, it succeeded on challenging targets like TNFα (where expert teams had failed) and outperformed a public competition's results on RBX1. While promising for accelerating therapeutic discovery, Anthropic has restricted such capabilities in its public model due to dual-use concerns. The prompts, data, and all designs are publicly released.

marsbit08/19 11:40

Claude Starts Independently Designing New Proteins, Efficiency Surpasses Human Experts by Dozens of Times

marsbit08/19 11:40

Sam Altman Names Him: The Most Important Researcher in AI, But Almost No One Knows Him

In a recent interview, Sam Altman gave a rare and high praise, calling Alec Radford "perhaps the most important, yet least known, researcher in AI history." Widely regarded as the true father of GPT, Radford is the lead author of foundational papers including GPT-1, GPT-2, CLIP, and Whisper, and contributed significantly to GPT-3, DALL·E, Scaling Laws, and GPT-4. Despite this monumental impact, Radford remains highly private, holds no PhD, and rarely gives interviews. Altman credits OpenAI's rise to hiring the then-23-year-old in 2016. Radford's early experiments, like training a model on Amazon reviews, led to the discovery of the "unsupervised sentiment neuron," revealing that models can learn unintended capabilities from simple next-token prediction. His pivotal move was applying the Transformer architecture to language modeling, creating GPT-1 in 2018. This established the "scaling" direction—focusing on increasing model size, data, and compute—which became OpenAI's core strategy, leading to GPT-2, GPT-3, and beyond. Radford later applied the same principles beyond text. His early work on DCGAN laid groundwork for image generation. At OpenAI, he contributed to Image GPT, DALL·E, and CLIP, demonstrating that a simple, scalable training task (like matching images to text) could yield powerful, general capabilities. His work on Whisper applied this to robust speech recognition. Described by colleagues as a "once-in-a-generation genius" and exceptionally kind, Radford is known for his low profile. In late 2024, he left OpenAI for independent research. His latest project, "Talkie," is a 13-billion parameter language model trained *only* on texts published before 1931. This experiment tests if a model with no modern knowledge can quickly learn new skills (like basic Python) from few examples, probing the boundary between memorization and true learning. True to form, as the world catches up, Radford is likely already working on the next big question.

marsbit08/17 08:13

Sam Altman Names Him: The Most Important Researcher in AI, But Almost No One Knows Him

marsbit08/17 08:13

Oxford and NUS Propose the Next-Generation World Model Direction: The Mental World Model is Here

Researchers from Oxford University and NUS propose Mental World Modeling (MWM), a framework that integrates mental variables like beliefs, goals, and social norms into world state representations, alongside physical variables. They argue that conventional world models, focused solely on physical dynamics, often produce predictions that are physically plausible but behaviorally incorrect in human-centric scenarios, as actions depend on both external and internal mental-social states. To validate MWM, they introduce MENTIS, a modular, training-free benchmark system that forces LLMs to perform explicit reasoning steps: parsing a scene into a coupled physical-mental state, generating agent-specific observations, simulating the joint state transition for candidate actions, and evaluating outcomes. Experiments on 8 LLM-based models show that explicit MWM reasoning significantly improves action prediction accuracy (F1 score rising from 63.3 to 87.9 on average), with the largest gains in social interaction scenarios. A key finding is that the primary remaining bottleneck is the simulation of state transitions. The work formalizes MWM and demonstrates its necessity for systems where behavior depends on variables not inferable from physical scenes alone.

marsbit08/13 04:46

Oxford and NUS Propose the Next-Generation World Model Direction: The Mental World Model is Here

marsbit08/13 04:46

OpenAI Researcher: We Don't Read Papers Anymore

An OpenAI researcher's remark that top AI labs "no longer read papers" sparked widespread discussion, highlighting a deepening crisis of trust in academic publishing. This sentiment followed exposure of questionable practices in an ICLR paper, where exceptional results were linked to undisclosed "tricks." A large-scale "experimental review" by SAI of 168 Oral papers from ICML 2026 revealed severe reproducibility issues. Of the 105 papers fully replicated, only 8 successfully verified over 80% of their claims, with a median verification rate of just 28-30%. Common problems included non-runnable code, missing files, incomplete documentation, and results mismatching those reported. Some papers even relied on now-offline models, making verification impossible. Specific cases involved an 8x inflation in claimed trained parameters and missing evaluation models from released code. Verifying a single ICML Oral paper had a median cost of around $8,900, with 17 exceeding $100,000. This creates a perverse incentive: flawed research carries high rewards (citations, jobs) with minimal risk of exposure, as verification is prohibitively expensive or impossible without code. While industry researchers at well-resourced labs may rely less on papers due to internal experiments and resources, academic and early-career researchers remain heavily dependent on publications for PhD applications, faculty positions, and entry into top labs. This creates a paradoxical system where papers are increasingly distrusted as reliable knowledge sources yet retain their gatekeeping value in career advancement. The situation underscores a critical need for systemic reforms to ensure scientific integrity and reproducibility in AI research.

marsbit08/09 23:31

OpenAI Researcher: We Don't Read Papers Anymore

marsbit08/09 23:31

China's 'Bio DeepSeek' Emerges: 4 Oxford Prodigies Let AI Take Over Life Science

China's 'Biology DeepSeek' Emerges: Four Oxford Alumni Aim to Let AI Take Over Life Sciences Following DeepSeek-V4-Flash's global impact, a Chinese counterpart for life sciences has arrived. Jindu Bio, founded by four Oxford University alumni, has developed GeneLLM, a multi-omics large language model. Published in top journals *Nature Communications* and *Advanced Science*, GeneLLM is the first model pre-trained directly on raw omics data, aiming to understand the "language" and "system" of life. GeneLLM treats the four RNA bases (A, U, G, C) as fundamental tokens, learning from raw sequencing data without relying on pre-defined annotations. It uses a Transformer architecture to predict the next base, processing trillions of RNA reads. With versions ranging from 1.5 billion to 30 billion parameters, it achieves high accuracy in disease prediction with significantly lower-cost, shallow-depth sequencing, making precision medicine more accessible. Beyond the model, Jindu Bio is building BioFord Harness, an infrastructure to connect AI with physical labs. This system translates scientific intent into executable commands for various lab equipment, manages scheduling, and creates a data feedback loop. Its platform features five collaborative AI agents for literature review, experimental design, scientific reasoning, lab scheduling, and data analysis, drastically speeding up research cycles. Crucially, it turns all experimental data—including failures—into valuable learning material. The founding team combines expertise in bioengineering, AI, computational biology, and business. After initial challenges in securing funding, the company completed four financing rounds in 2023 following China's national push for "AI+" initiatives, supported by prominent investors like Sequoia Capital China and Gaotegaj Investment. In the broader AI for BioScience landscape, Jindu carves a unique niche. Unlike digital-only AI scientists or capital-intensive fully automated labs, it focuses on the "last mile" of infrastructure—orchestrating the entire research workflow by bridging AI models with existing laboratory hardware. Its long-term vision is to build an intelligent operating system for life sciences, defining a new, scalable paradigm for scientific discovery.

marsbit08/04 13:56

China's 'Bio DeepSeek' Emerges: 4 Oxford Prodigies Let AI Take Over Life Science

marsbit08/04 13:56

Open-source Claude Science is here: zero dependencies, MIT license, built-in 30+ research Skills

OpenAI4S (Open AI for Scientist), an open-source, MIT-licensed research agent, has been released as a public replication of the closed-source Claude Science. Developed by the Peking University & Yuankong AI Agent Joint Laboratory, it adopts a "Code-as-Action" approach, where the AI directly generates and executes Python/R code within a persistent kernel, enabling complex, multi-step scientific workflows like data retrieval, analysis, visualization, and report generation. The platform functions as a collaborative research workbench. It comes pre-built with over 30 scientific "Skills" for tasks such as protein structure analysis, molecular docking, and literature review—14 of which require GPU/professional compute. A key principle is its strict "no-fabrication policy": it uses real data from sources like UniProt and RCSB, performs actual computations (e.g., protein folding on GPU servers), and reports errors rather than generating simulated results. OpenAI4S supports connecting to users' own computing environments (e.g., GPU clusters) to handle demanding tasks while maintaining workflow integration. The project is open-sourced to encourage community contribution, inviting researchers and developers to expand its Skills, core architecture, and applications across diverse scientific fields.

marsbit08/04 10:22

Open-source Claude Science is here: zero dependencies, MIT license, built-in 30+ research Skills

marsbit08/04 10:22

10,000 Scientists Get 1 Year of Free Access: OpenAI Brings the Scientific Research Pipeline into ChatGPT

OpenAI has launched the "ChatGPT for Academic Researchers" program, offering free one-year access to its flagship models for 100,000 university researchers globally, with 10,000 spots available this summer. Selected institutions include prestigious centers like ENS Paris and the IAS at Princeton. The initiative provides an integrated research workspace within ChatGPT, bundling tools like ChatGPT, ChatGPT Work, and Codex, along with expanded Deep Research capabilities, higher usage limits, and specialized tools for life sciences. The suite connects to platforms like Zotero and GitHub, aiming to streamline the entire research workflow from literature review and coding to data analysis and manuscript drafting. OpenAI notes that about 1.3 million people already use ChatGPT weekly for advanced science and math. The program targets building long-term user dependency by embedding these tools into daily research habits. However, access comes with limitations: it does not include API credits or model weights, and eligibility is restricted to verified academic researchers from supported countries. This approach contrasts with Anthropic's "AI for Science" program, which offers API credits but not an integrated workspace. Both companies emphasize preventing misuse by withholding model weights, a point of contention for AI researchers seeking transparency. The core strategy remains clear: provide a powerful, integrated environment to foster user reliance ahead of the post-free period.

marsbit07/30 11:51

10,000 Scientists Get 1 Year of Free Access: OpenAI Brings the Scientific Research Pipeline into ChatGPT

marsbit07/30 11:51

He Let GPT-5.6 Sol Run for 33 Hours Straight to Tackle Fermat's Last Theorem, Forcibly Terminated by the System

This article discusses a real-world experiment by expert Michael P. Frank to test if an AI, specifically GPT-5.6 Sol, could autonomously make progress on a major unsolved mathematical problem: finding a simpler proof for Fermat's Last Theorem. The AI was tasked with exploring specific mathematical pathways and maintaining rigorous notes over approximately 33 hours. However, the session was terminated by OpenAI's systems. The AI itself suggested two possible reasons for the stoppage: excessive resource consumption, or OpenAI having previously failed on similar problems and wishing to conserve computational resources. The AI reported its work primarily involved refining plausible ideas into precise, verifiable statements, most of which were subsequently disproven or excluded—effectively creating a map of dead ends rather than a proof. The incident sparked debate online. Some speculated that OpenAI might deliberately restrict public access to its most powerful models to maintain a competitive edge or avoid regulatory scrutiny, rather than allowing users to potentially solve landmark problems. OpenAI researcher Noam Brown countered this, arguing that a user solving a major problem would be tremendous publicity. Others offered technical explanations, suggesting the termination could be due to standard safety mechanisms preventing infinite loops, or even a known bug in the GPT-5.6 Sol version that disrupts long-running sessions. The story highlights the practical challenges, technical limits, and broader strategic questions surrounding the use of advanced AI for open-ended, high-stakes research.

marsbit07/28 12:26

He Let GPT-5.6 Sol Run for 33 Hours Straight to Tackle Fermat's Last Theorem, Forcibly Terminated by the System

marsbit07/28 12:26

AGI Countdown: OpenAI's Chief Research Officer Makes Major Statement — The Window for Humanity is 'Very Small'

The countdown to AGI has begun, according to OpenAI's Chief Scientist Mark Chen, who states the window for human-centric progress is "very small." Chen argues that AI is reaching a point where models can perform "self-sustaining research," autonomously driving innovation in fields from mathematics to programming. He points to the proliferation of AI's "superhuman" insights—akin to AlphaGo's legendary "Move 37"—across disciplines as evidence of this shift. Chen firmly dismisses claims that scaling laws are plateauing or that pre-training is dead, asserting the field remains on an exponential curve. He cites OpenAI's successful bet on reasoning models like o1 as proof that fundamental breakthroughs are still possible. The future of research, he suggests, lies with "Vibe Researchers"—humans who provide high-level direction and "taste" while AI handles execution and orchestration of complex, long-horizon tasks. However, significant hurdles remain. Chen highlights a "benchmarking crisis," where models can overfit to existing tests without gaining true generalization. He also notes the "jagged frontier" of AI capabilities, where systems excel at advanced reasoning but struggle with contextual, continual learning from everyday experiences. Despite these challenges, he expresses confidence that these gaps will be closed. In a personal reflection, Chen shares that post-AGI, his wish is to open a noodle shop—a metaphor emphasizing that when AI masters knowledge and innovation, uniquely human experiences, warmth, and storytelling will become the ultimate form of value.

marsbit06/30 08:37

AGI Countdown: OpenAI's Chief Research Officer Makes Major Statement — The Window for Humanity is 'Very Small'

marsbit06/30 08:37

活动图片

Othersright-arrow

Regulatory Policyright-arrow