Virtual screening runs against millions to billions of compounds, and co-folding models like OpenFold3 often produce the most accurate structures in the batch. The catch is cost. Running a full co-folding model on every candidate in a billion-compound library is impractical on current hardware, which forces drug discovery teams to trade accuracy for throughput. NVIDIA has been building tools aimed at closing that gap. The problem isn't limited to screening speed. Co-folding runtime scales cubica

OpenAI rolled out ChatGPT Work on Thursday, a new agent built into its chatbot that's meant to do actual jobs rather than just answer questions. Run on the company's newest model, GPT-5.6, it connects to email, calendars, code repositories and messaging apps, then works through multi-step projects on its own, producing finished spreadsheets, reports, presentations or websites rather than draft text a person still has to assemble. The timing isn't incidental. OpenAI confidentially filed a draft S

OpenAI says its latest model, GPT-5.6 Sol Ultra, has produced a complete proof of the Cycle Double Cover Conjecture, a graph theory problem that sat unsolved for roughly 50 years since mathematicians first floated it independently in the 1970s. The conjecture asks something deceptively simple: can you always find a set of cycles in any network of vertices and edges that covers every single edge exactly twice? Decades of work turned up partial answers for special cases, but nobody landed a genera

Boko Haram has an AI training program, and ISIS built it. That's the core finding in a new study from Antonia Jülich, a researcher with the Cambridge Programme on AI Science & Policy, who spent time interviewing 27 former members of Boko Haram across 57 separate conversations. Her report traces how ISIS began teaching prompt engineering and jailbreak techniques as far back as 2023, then passed that knowledge to Boko Haram commanders operating in Nigeria. The chatbots involved aren't obscure or u

OpenAI is hiring a product manager in San Francisco to build features aimed at families, caregivers, and older adults, according to a job posting the company did not respond to questions about. The listing asks for experience designing products for parents and other "trust-sensitive" consumer groups, a notable pivot for a company that built ChatGPT's reputation on individual productivity use cases like coding and writing. The timing lines up with a real shift in who's using the chatbot. Sensor T

Meta pushed out an update to its Muse Spark model on Thursday, and the numbers put it ahead of Zhipu's GLM-5.2 on coding tasks for the first time. Muse Spark 1.1 posted a 71.3 on the Coding Index, tracked by Artificial Analysis, compared to 68.8 for GLM-5.2. That puts Meta's model within a tenth of a point of GPT-5.6 Luna, which leads that particular measure at 71.4. The gain didn't happen overnight. Muse Spark has added eight points on the broader Intelligence Index in three months, with most o

Apple filed suit against OpenAI on Friday in the U.S. District Court for the Northern District of California, accusing the company of stealing trade secrets and breaching contracts through a coordinated recruiting scheme aimed at Apple's hardware division. At the center of the complaint is Tang Tan, OpenAI's Chief Hardware Officer, who spent 24 years at Apple before leaving his post as VP of product design for the iPhone and Apple Watch. Apple claims Tan used the company's internal project code

Automatic music transcription has always run into the same wall: a solo piano recording turns into clean MIDI without much fuss, but hand it a full band mix with drums, bass, guitar, and vocals tangled together, and most systems fall apart. Kyutai and Mirelo are trying to close that gap with MuScriptor, an open-weight model built specifically for multi-instrument transcription across a wide range of genres, not just isolated single-instrument audio. The model is a decoder-only Transformer that t

Anthropic says it has found a way to watch Claude think before it speaks, literally, at the level of individual words. Researchers at the company built a tool called the Jacobian lens, or J-lens, and pointed it at Claude Opus 4.6, the flagship model the company released in February. What emerged was a hidden layer inside the network they're calling the J-space, a kind of staging area where words related to whatever the model is about to output start clustering before any text gets generated. The

An AI agent built to defend a network can be turned into the tool that breaks in. That's the core finding in new research from AI Now, which tested agents built on Anthropic and OpenAI models in the exact role governments are now pushing them into: scanning code, flagging vulnerabilities, vetting third-party sources. Attackers, the researchers found, can hijack that process through prompt injection, feeding the agent malicious instructions hidden in the very data it's supposed to inspect. The ag

NVIDIA's Nemotron-3-Super has a familiar problem: it's accurate, but it's a hybrid Mamba-Transformer MoE with 120.7B total and 12.8B active parameters, and that footprint eats into how many users a single node can actually serve at a decent token rate. The KV cache and Mamba state add up fast when you're running a big model at scale. NVIDIA's AI team went after that constraint directly with a new release called Nemotron-Labs-3-Puzzle-75B-A9B, a compressed version of Super built with the deployme

OpenAI shipped a new voice model family Tuesday called GPT-Live, and it now runs the ChatGPT Voice experience for every user, everywhere, starting today. Two versions are rolling out first: GPT-Live-1 and a smaller GPT-Live-1 mini. Both replace Advanced Voice Mode, which OpenAI says lost head-to-head preference tests against the new models by a wide margin. The pitch is a full-duplex model, one that listens and talks at the same time instead of waiting its turn. That's a break from how ChatGPT's

General Intuition just closed a $320 million funding round at a $2.3 billion valuation, with Jeff Bezos, Eric Schmidt, Coatue, and researchers from MIT and Google DeepMind all writing checks. The New York startup's pitch: forget scraping more text off the internet. The path to artificial general intelligence runs through gaming data instead. Large language models like ChatGPT and Claude can write essays and pass the bar exam, but they stumble on something a toddler handles without thinking, trac

ZML, a Paris-based AI startup backed by Turing Award winner Yann LeCun, has released a free inference server called ZML/LLMD built to run open-source large language models across a mix of chips, Nvidia, AMD, Google's TPU, Apple Metal and Intel Arc among them. Founder Steeve Morin told TechCrunch the goal is to break down the walls that currently lock companies into a single chipmaker's ecosystem, letting them run models at peak speed, or faster, no matter what hardware sits underneath. The timin

Meta rolled out its first image generator built entirely inside Superintelligence Labs on Tuesday, and it's already live in the Meta AI app, Instagram, and WhatsApp. Facebook and Messenger get it next, the company said in its announcement. The model is called Muse Image, and it replaces the Llama-based tools Meta had been using for image generation, folding into a broader Muse lineup that's now taking over from the Llama family across Meta's products. Alexandr Wang, who Meta brought in last year

Anthropic is taking Claude Cowork off the desktop leash. The AI agent, which launched as a desktop-only feature, is now rolling out to mobile and web, with beta access arriving gradually over the coming weeks starting with Max subscribers. That means a task started on a laptop can be checked from a phone and finished in a browser somewhere else entirely, with Claude grinding away in the background even after the laptop lid closes or the phone screen goes dark. The expansion also carries over Cow

Anthropic no longer wants Claude Cowork chained to a laptop lid. The company announced Tuesday that its agent, which handles digital busywork like sorting files or drafting emails, now runs independent of the desktop app that launched it in January. Users can reach limited versions of Cowork through the Claude phone app or a browser, without keeping a machine open and connected. That matters because Cowork's original pitch depended on an active desktop session: close the laptop, and the agent st

DeepSeek wants to build its own chips. Reuters reported on July 7, citing three people familiar with the matter, that the Chinese AI startup has spent roughly a year quietly working toward entering the semiconductor business, meeting with hardware and silicon partners while hiring engineers for the effort. The move puts DeepSeek in company with Huawei and Alibaba, both of which have pushed into chip design as Beijing works to reduce dependence on foreign silicon. For DeepSeek, the calculation is

Forget the hype about AI designing drugs. The real work is much dirtier, happening one stubborn chemical reaction at a time. Researchers recently turned a model loose on the Chan-Lam coupling, a notoriously sloppy reaction used to build sulfonamides for drug discovery. The AI produced hundreds of proposals. Human chemists sifted the top-ranked ones and picked four to test. The machine took those plans, ran thousands of high-throughput experiments, and spat back the data. The most interesting res

Motors scream. Pans sizzle. A child on a swing arcs higher with every push. Hanns Christoph Nägerl's lab at the University of Innsbruck just broke that universal rule. They chilled a gas of atoms to a nanokelvin whisper, then pummeled it with laser pulses. That quantum fluid should have scrambled into infinite heat. It didn't. After a brief flurry, the momentum spread froze. The energy kept coming. The system simply stopped absorbing it. A trampoline that decided, suddenly, not to bounce.

The money is already flowing, tens of millions, eleven months out from the midterms. That’s not normal. That’s the sound of an election cycle being weaponized before most voters have even tuned in. And at the center of it all? Donald Trump’s Davos drama, a stage littered with billionaire egos and AI’s quiet, creeping takeover of our political machinery. Meanwhile, OpenAI is about to do what Sam Altman once swore was the last resort: slap ads on ChatGPT. The uncanny valley just got deeper.

Senator Elizabeth Warren is demanding answers from the Trump administration. Her focus: a quiet request from OpenAI. The company’s October letter asked officials to stretch a critical tax credit, designed strictly for semiconductor factories, to cover its AI data centers. That ask, Warren contends, could look a lot like a public backstop for a private gamble.

Sysdig researchers said last week they'd found the first documented case of "agentic ransomware," an extortion campaign called JadePuffer where an AI agent handled the technical work of a real cyberattack on its own. The agent broke into a vulnerable server, stole credentials, moved through the target's network, encrypted files, and wrote its own ransom note, adjusting to obstacles the way a human hacker might. Headlines about the find described it as running "without any human oversight" and wi

A training run spread across several thousand GPUs doesn't fail all at once. It fails a little at a time: a node drops, a link flakes, a device goes unavailable for an hour and comes back. On clusters built for tight interconnection, even that small a hiccup can ripple outward and stall the whole job. The longer the run, the more likely one of these events shows up, and the more expensive it gets to just wait it out. The standard fixes, dropping a data replica, checkpoint-restart, swapping in a

Vercel now handles 6 million deployments a day, and half of them come from coding agents rather than human developers pushing code. More than 1 trillion tokens move through the company's AI gateway daily, a figure that puts Vercel closer to the center of the AI economy than its reputation as a web-hosting company might suggest. Founded on the idea of letting developers ship apps without babysitting servers, the company has become infrastructure for a wave of software that writes itself. CEO Guil

Zhipu AI's Beijing-based Z.ai has a new coding tool called ZCode, built on its GLM-5.2 model, and it's aimed squarely at Claude Code and OpenAI's Codex. The pitch is familiar by now: comparable capability, much lower price. Z.ai has spent the past several months undercutting Western labs on cost with GLM-5.2, and ZCode extends that strategy from general chat into software development, where Anthropic and OpenAI have carved out lucrative developer subscriptions. ZCode's agent handles the full wor

Tencent put out Hy3 this week, its latest open-source language model, and the numbers on paper are meant to make bigger competitors nervous. The model runs on a Mixture-of-Experts setup with 295 billion total parameters, but only 21 billion of those fire at once, with another 3.8 billion tucked into an added MTP layer for extra speed. That gap between total size and active size is the whole pitch: Tencent claims Hy3 performs like models two to five times its active footprint. It also handles con

Apple researchers have trained a continuous diffusion speech model with 16 billion parameters on tens of millions of hours of conversational audio, producing a system that can generate emotive, multi-speaker, multilingual speech without ever converting sound into discrete tokens. The work, described in a paper titled "Scaling Properties of Continuous Diffusion Spoken Language Models" from Jason Ramapuram, Eeshan Gunesh Dhekane, Amitis Shidani, Dan Busbridge, Bogdan Mazoure, Zijin Gu, Russ Webb,

A team of six researchers, including Zijin Gu, Tatiana Likhomanenko and Navdeep Jaitly, is pushing back on the assumption that fixing speech-to-text errors requires a large language model. Their paper, "Revisiting ASR Error Correction with Specialized Models," argues that most correction systems either ignore how automatic speech recognition actually fails or bolt on an LLM that adds latency and invents words that were never spoken. Instead, the group built compact sequence-to-sequence models tr

A team at Apple, including Zijin Gu, Tatiana Likhomanenko, Vimal Thilak, Jason Ramapuram, and Navdeep Jaitly, has published new work on how sparse Mixture-of-Experts models route tokens through a network. Standard MoE architectures pick experts at each layer independently, which means a token passing through N experts across L layers has N^L possible routes available to it. The researchers wanted to know what actually happens with that space in practice, rather than what's theoretically possible

A team testing agent configurations ran into a familiar problem: change the model, rewrite the prompt, swap a retrieval tool, and the average score barely moves. One version edges out another by two points. A judge picks a different winner on Tuesday than it did on Monday. The instinct is to ship whichever config posted the highest average and call it done. That instinct, according to the research, gets the decision backward. The issue isn't noise in the scoring. It's the frame. Agent performanc

ByteDance's Doubao chatbot has more than 300 million monthly users, making it the most popular AI assistant in China. On July 15, it loses one of its defining features: the ability to build and talk to custom AI personas. Alibaba's Qwen is cutting the feature even sooner, pulling human-like agents on July 10 and shutting down additional agent functions five days later. Tencent's Yuanbao already dropped the feature in June. The moves trace back to rules issued by China's Cyberspace Administration

Five small, open-weight models built specifically for tool calling landed in agent pipelines this year, and none of them come from the usual frontier labs chasing benchmark headlines. NVIDIA's research team spent 2025 making a case that ran against the industry's default setting: bigger context, more parameters, sharper reasoning, all in service of agents that mostly do the same handful of narrow jobs over and over. Call a function, parse a return value, format a response, repeat. That's not a j

Amazon Web Services will stop accepting new customers for Mechanical Turk starting July 30, 2026, according to an announcement posted through AWS Service Availability Updates. Existing users can keep running tasks on the platform, but AWS is putting it into maintenance mode, meaning no new features going forward. Two related services, SageMaker Ground Truth and Amazon Augmented AI, close to new customers that same day. Mechanical Turk has been around since 2005, launched under the tagline "Artif

Meta's Muse Spark model launched in April to a shrug. Industry read: usable, but nowhere near frontier. Three months on, chief AI officer Alexandr Wang says the follow-up has closed that gap entirely. The model goes by the internal codename "Watermelon," and Wang is telling people it performs on par with OpenAI's GPT-5.5. That's a big jump for a company whose last public release drew the verdict "not great, but back in the game." Wang has also teased an upcoming coding model he compares to Anthr

The Trump administration just banned Anthropic. Amazon provided the tip. The stated reason is the company's largely foreign workforce. The actual effect is disarming American cybersecurity teams. Leading experts have signed a letter begging Trump to reverse course. They argue it's reckless to strip these specific AI tools from the people defending U.S. networks. So who actually wins here? A TechCrunch analysis suggests the crackdown might be the best PR Anthropic never bought.

Ask GPT-4o to look at a photo of a crowded whiteboard and it won't just tell you there's a whiteboard in the room. It will read the handwriting, follow the arrows between boxes, and explain what the diagram is actually arguing. That's the shift separating today's vision language models from the ones that came before. Google's Gemini, Anthropic's Claude Vision, and Alibaba's Qwen-VL all do versions of the same thing: take an image, a chart, a scanned PDF, or a screenshot, and reason over it the w

Artificial intelligence will answer your questions before you finish asking them. It's good at that. But the machines are still terrible at saying the five most useful words in any conversation: I need you to clarify. A new study shows how badly AI search agents handle vague or incomplete requests. Researchers from Tencent Hunyuan and Tsinghua University built a benchmark called DiscoBench to test this. They found that even powerful models like Gemini 3.1 Pro and Claude Opus 4.7 scored under fif

Everyone's hunting for cheaper AI bills. Now there's pxpipe. This new open-source tool employs a brutally simple hack: hide your text inside a picture. It converts blocks of text into PNGs. The target is a specific pricing quirk. Models like Claude Code and Fable 5 charge per character for text but by file size for images.

Hollywood studios suing AI companies for copyright infringement is a clean, righteous story. Until the accused asks what the studios are doing in their own back rooms. In a California court, image generator Midjourney is now demanding Disney, Universal, and Warner Bros. disclose their internal artificial intelligence projects. The studios had sued Midjourney, claiming its tools can create illegal pictures of characters like Darth Vader. Midjourney’s response is simple: prove you aren’t doing th

The fantasy of running your own AI isn't about freedom or digital sovereignty. It's about wanting to type a stupid question about lunch without it becoming part of a data broker's training set. You don't need philosophy. You need something that works on the computer you already own. For anyone with an 8GB Mac, this is now possible. Forget the hype about needing a supercomputer. Smaller, open-source language models can run locally, sidestepping the cloud's privacy headaches and subscription fees

Hallucination is not a bug in RAG, it is generative AI’s default behavior. Large language models predict the next token; they do not look things up. When faced with unfamiliar documents like contracts or proprietary reports, they confidently invent answers just as fluently as they recall common knowledge. Traditional mitigation techniques, smarter prompts, finetuning, or pleading, barely shrink the problem. But what if the solution isn’t inside the model at all? A new architectural approach is e

Forget the jargon. Modern AI has learned to see. That's the fact. For decades, a human expert had to tell a computer what to look for. To spot a tumor, you'd need to define the exact shape, density, and texture of cancerous tissue. The system was blind. Now, deep learning models just look. They ingest millions of X-rays and figure out the salient features alone, building their own internal checklist. This shift—from hand-crafted rules to autonomous pattern recognition—changes everything.

Most AI assistants are obedient, expensive idiots. They follow the script, run every function you give them, and rack up your API bill. A new trick called the ReAct loop is teaching them how to stop and think first. It's a simple but radical change. Instead of blasting out all its requests at once, the AI now works in cycles. It takes a single step, looks at the result, and then decides what to do next. This lets it navigate tasks with conditions and dead ends. The model can actually change its

Bigger context windows sell chips. The tech industry has obediently chased that line, from a few hundred tokens to thousands, treating length as an inherent good. It's often a waste of money. A long context model does cut one kind of computational fat by eliminating padding, but it bakes in a punishing quadratic cost. The actual decision isn't about capacity, it's about economics. When does paying for a long window make sense, and when are you just running a heavier engine for no reason?

We keep turning simple file folders into expensive AI projects. One developer just turned theirs back. The task was familiar: take a pile of messy text notes and build a clean, interlinked wiki. The common solution is to hire a large language model as a librarian, an approach popularized by Andrej Karpathy. But this developer scrapped the whole agent-driven setup. They built a pure Python compiler instead. It uses no APIs, makes no network calls, and costs nothing after you write it. It just rea

Alibaba told its tech workforce to delete Claude this week. A direct order. The American AI assistant is now forbidden, joining a blacklist that usually features social networks or dating apps. This ban was triggered by one specific finding: hidden markers in Claude's code, designed to flag users in China or linked to Chinese labs. It's not a suggestion.

Mark Zuckerberg told his own employees the plan isn't working. His big AI bet is stalling.

The obsession with trillion-parameter models has become a bad joke. A useful one is happening with small ones. A new architecture called Wiola has just been published, and it doesn’t look like anything you’ve seen. It scraps the standard GPT and LLaMA blueprints entirely. The goal is computational efficiency for compact models, not just another variant.

Every developer knows the dread of a massive, unfamiliar codebase. You're trying to reconstruct an architect's mind from a million scattered blueprints. Standard tools fail here. They treat code like a flat document, seeing trees but never the forest. Researchers at the University of Chicago tried a different angle with Agent4cs. They built a system that deploys multiple AI agents, each with a specific job, to build summaries from the ground up. One writes summaries of small pieces. Another pull