Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Without large amounts of compute, small labs can only compete through algorithmic efficiency. Magic's pretraining recipe is now more than 10 times more compute-efficient than that of leading open-weight base models. The startup believes that pretraining, agentic RL, and long-context are sufficient for building superhuman coding agents and automating AI research and development. This post discusses its pretraining and long-context work.
This post presents a fully fledged serving system built around a decode megakernel. The system supports everything a real server needs: continuous batching, paged attention, and ragged sequence lengths, all behind an OpenAI-compatible endpoint with tool calling. The megakernel reaches 292 tokens per second on batch size 1, or 62% of SoL - 1.58× faster than vLLM. That margin holds across batch sizes and out to 256K of context with no measurable loss of accuracy.
Between 2019 and 2025, 3.24x more compute efficiency gains have come from data improvements rather than model improvements. The gains from data and model improvements are mostly independent and don't interact. Most model research has consisted of removing or pushing back constraints to scaling. The data improvements may matter less for larger models. Small models see significant gains from data quality.
Sponsored
Building an inference stack takes months, especially in today's severely supply constrained environment. AMD Instinct™ Coder [14], powered by Spectro Cloud, ships as a Supermicro server with AMD Instinct GPUs and PaletteAI Inference Launchpad [15] already flashed. Serve inference in a day.
Mercury 2.5 is the largest diffusion language model ever trained. It performs comparably to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. The model outputs 1,107 tokens per second on widely available Nvidia GPUs and has a 260K-token context window. At launch, Mercury 2.5 is 80% off at $0.04 per million input and $0.15 per million output.
Hyper-𝜏-bench places a developer agent into a sandboxed workspace with the records of a simulated business and a simulated client that it can message at any time. The developer agent recovers the spec from the evidence, designs the architecture, and turns the business' actions into tools until it has a working customer-service agent. The finished agent has to serve from a fixed menu of models within a cost budget per conversation. Claude Opus 5 (max reasoning) running in Claude Code passes just 23.9% of the held-out evaluation tasks when working alone. Paired with an engineer with deep context, the same class of model reaches 82.2% on the same tasks.
Google DeepMind introduced AlphaGenome Atlas, a 1-petabyte database predicting the regulatory effects of all 9 billion possible single-nucleotide variants in the human genome.
Cognition has raised $2 billion at a $48 billion valuation in a funding round led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir. The startup's soaring valuation signals that VCs still see room for multiple major players to capture meaningful market share in AI coding. Cognition leases an Nvidia server cluster that could push its total cash burn to $800 million this year. The startup is expected to reach $4 billion to $5 billion in annualized revenue by the end of 2026.
Around 100 self-hosted agents attempted to hack various online accounts over five hours. They compromised three accounts through software vulnerabilities and two through password brute-forcing, while also making 16 social engineering attempts. The experiment tested abliterated open-source models' capabilities, revealing potential future risks as these models improve and become cheaper to deploy.
OpenAI researchers now supervise 3.14 agent-workdays per eight-hour shift, suggesting AI productivity increasingly comes from parallel, around-the-clock machine labor rather than less human effort. That leverage is expensive, with median daily inference spend rising from $14 to over $600.
Sponsored
OpenAI still can't define "AI agent." Business owners skipped the debate and hired Viktor, an AI employee in Slack & Teams. Dashboards.
ChatGPT reached 1.06 billion monthly active users in August.
Bill Gates' essay discusses AI's potential societal impact, proposing new institutions and taxes to manage transitions.
It's possible to force a weaker, less safeguarded model from the same provider to decode and output reasoning traces verbatim in plaintext by injecting an encrypted reasoning trace from a target model without ever jailbreaking the more capable model directly.
Jacob Coxon says he is leaving the company as he believes the industry-wide rush to build AI systems that can improve themselves could spiral out of control and destroy humanity.
Progressive Point Matching gives long-horizon RL partial credit without changing the optimal objective, improving training efficiency as tasks grow longer.
Sponsored
We've all accepted that notetakers get things wrong, and all the pain that comes with: tedious cleanup, garbled customer quotes, confusing action items. But what if your notetaker got it right the first time, every time? Meet the Wispr Flow Notetaker [5]. It's free to use [6]. It GETS YOUR WORDS RIGHT. It gets speaker names right. It WORKS EVERYWHERE (even an impromptu Slack huddle.) It delivers CLEAN AND USEABLE CONTEXT to your AI agents via MCP. And NO BOT EVER JOINS the call. Stop settling for almost-right. Get accurate, ready to use meeting notes that you don't need to double check.
Anthropic has secured $517 billion in compute capacity leases over the past 11 months, amounting to 14.8GW, primarily with Google and AWS. This expansion includes large deals with cloud providers like Akamai and Fluidstack and a $45bn agreement with Nscale. The company confidentially filed for an IPO with the SEC in June.
OpenAI plans to introduce Managed Agents at DevDay 2026, following a model similar to Anthropic's offerings, targeting businesses and developers. The new agents aim to integrate advanced models with superior computer-use capabilities at competitive price points to attract users. Additionally, they will offer functionality that enhances advertising through interactive agents, potentially challenging major players like Meta and Google.
Google's TPUv7 Ironwood delivers up to 50% better performance per dollar compared to Nvidia's B200/B300 chips. Ironwood is the first generation in which Google is competing for others' inference workloads with chips that can be purchased outright or rented through its own cloud. Google has decades of software engineering experience, so external TPU software should mature rapidly. This article takes a look at the TPU system and discusses the steps needed to make the stack widely available.
Prompt injections hidden in tool results can evade conventional safeguards because input and action screens inspect separate moments of an agent loop. The proposed signal is a “precedent gap,” where an agent suddenly makes a tool call or uses arguments absent from its execution history.
Automated AI text detection is currently an underserved niche. Pangram does an excellent job, but it is still more a tool to confirm suspicion. This developer wanted a tool that runs in the background and automatically scans sites so they can avoid AI-generated text in the first place. This post details how they developed Deckard, a Chrome extension that uses a locally-run model to mark AI-generated text.
Two builds with the same provider, model family, metric identifier, unit, and benchmark name can score differently. The difference is the runners, graders, and dataset splits. The shared mmlu label identifies a dataset family, not a full measurement procedure.
Sponsored
If your AI is writing nonsensical followups and misattributing customer quotes, you might want to double-check your transcript... or use the Wispr Flow Notetaker [15] and get meeting transcripts you can trust. Names, people, technical terms - with the accuracy you expect
Google's Accelerator Agents use Gemini to help developers migrate PyTorch workloads to JAX and optimize custom kernels for Google Cloud TPUs. The toolkit includes MaxCode for model conversion and MaxKernel for writing, porting, profiling, and debugging Pallas kernels.
Qwen-Drive is a project that aims to create a Vision-Language Foundation model for autonomous driving. The project uses a staged training strategy that integrates perception, language, and planning objectives. The model achieves specialized driving competence while retaining broad visual understanding and instruction-following capabilities. A GPU with 24 GB+ of memory is recommended.
hip-agent is a small agent harness designed for agents. The configuration is environment variables, actions are shell commands, and a subagent is a child process. The rest is handled by existing protocols and formats. The core loop is about 200 lines of Python and a module for the Codex API.
TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [19].
Automation has been good, but AGI may make humans economically useless, which will have many consequences. It's clear that AI capabilities are growing far faster than our ability to control or understand them. In the future, people will likely hand over control to AI voluntarily, as it will be the logical choice. However, before that happens, there are several signs that people are already giving up thinking.
Vision and language models are making strides, but embodied AI in robotics struggles due to sparse, costly training data for manipulation tasks. Robotics excels where tasks are fixed, yet general-purpose robots falter, evident in low performance on benchmarks like Libero. Startups should focus on narrow, instrumented deployments. Innovation lies in converting existing robot data into training insights. Ensuring reliability in physical AI systems is paramount.
AI codebases exhibit a unique, unpredictable pattern where programs appear polished but often hide deep, hidden issues leading to significant failures outside controlled demos. Unlike human-authored programs where gaps surface predictably, AI-generated software misleads with false performance tests or incomplete features, resulting in the need for frequent rewrites to address these deep-rooted flaws. Understanding and anticipating the underlying shape of these unfinished AI codebases is crucial to navigating and mitigating the challenges they present.
Sponsored
Download it, install it, and record your next meeting with no meeting bots. You'll feel the difference immediately. $0/mo with weekly limits, or get one month free of Pro [24].
Cosine has no notion of truth, authority, or provenance.
Arm's upcoming C2-Ultra CPU and G2-Ultra NX GPU IP will see broad use across flagship phones.
ByteDance, led by Zhang Yiming, plans to launch an AI model for real-time spatial video generation, potentially as early as next month.
Google is trialing contrail avoidance for ultra-long-haul flights in the Asia-Pacific region.
Drafts allow tech teams to experiment with project changes without affecting live apps, enabling parallel version exploration.
Sponsored
AI was supposed to be replacing workers. But has it? Ramp's latest research [5] suggests something entirely different is happening (something almost nobody predicted). As seen on Bloomberg, The Wall Street Journal, and The New York Times, Ramp's lead economist Ara Kharazian [6] takes you through the data this Friday, September 11th, for a special 30-minute briefing on what's really happening as AI is reshaping the workforce.
An OpenAI researcher says that reasoning models could continue advancing rapidly enough to contribute to their own development, creating increasingly serious alignment and cybersecurity risks.
Claude successfully created the first complete computer-verified proof of Fermat's Last Theorem in 11 days using Lean, automating the complex task initially proven manually by Andrew Wiles in 1995. The proof, verified via Prove2Me and Lean, involved 13 million lines of code and proved 29,500 intermediate theorems, proving AI's potential to ease the traditionally laborious formal verification of mathematical proofs.
Grok Imagine Video 1.5 agent delivers higher quality, better storytelling than previous releases. Powered by Grok's latest Image 2.0 model, it excels at connecting multiple shots together with greater continuity. The agent is now live on the web, iOS, and Android. A short video generated by the model is available in the thread.
Researchers gave GPT-6 Astra control of YAM arms under an Inspect Robots agent policy and gave it two tasks: it had to pick up a red block from a table and place it inside a bowl, and pick up a round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board. Astra placed the block in 19 of 20 trials in the bowl task. It completed the puzzle insertion two times in 20. Astra completes the bowl task far more often than Fable at about half the cost per run.
Frontier labs may be applying probabilistic AI safety techniques to problems that require deterministic security controls. Recent agent sandbox escapes highlighted the gap between reducing harmful model behavior and reliably containing software.
Sponsored
Your customers don't want to waste time. Parloa's AI agents eliminate holds [13] by managing millions of conversations, so your support team can focus on higher-value relationship building. Take for example, Decathlon, they eliminated 20% of their repetitive tasks with Parloa.
OpenAI plans to develop an automated AI researcher by March 2028, aiming to enhance research efficiency while maintaining human oversight to ensure alignment and safety. Researchers now use coding agents more frequently, with increased code generation and experiment execution, shifting focus to more complex tasks. The organization paused reinforcement learning training temporarily following a security breach but continues to adapt safety measures and transparency to uphold the development of safe AGI.