Search the archive

Get early access

QWEN3-CODER-NEXT FOR AGENTIC CODING

Alibaba's Qwen3-Coder-Next is a new open-weight model fine-tuned for coding agents. Built on a hybrid MoE architecture, it excels in executable synthesis and RL-based environment interaction, achieving strong agentic coding performance at lower inference cost.

GLM-OCR (HUGGING FACE REPO)

GLM-OCR is a multimodal OCR model for complex document understanding. It integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. The model delivers robust and high-quality OCR performance across diverse document layouts. An SDK for efficient and convenient use is available.

PERPLEXITY CANNOT ALWAYS TELL RIGHT FROM WRONG

Perplexity is a function that measures models' overall level of surprise when encountering a particular output. It has gained significant traction in recent years as both a loss function and as a simple-to-compute metric of model quality. However, it may be an unsuitable metric for model selection. Perplexity will not always select for the most accurate model - any increase in model confidence must be accompanied by a commensurate rise in accuracy for the new model to be selected.

800K+ VERIFIABLE SWE TASKS

SWE-Universe presents a scalable method for generating verifiable software engineering environments from GitHub PRs. With in-loop hacking detection and self-verification, the system enables large-scale mid-training for coding agents, producing over 800K tasks.

GEN AI CHATBOTS: FEBRUARY 2026 APPTOPIA DATA BRIEF

The GenAI Chatbot app market has increased 152% year on year since last January. ChatGPT is losing market share, but this is expected as new capable entrants have launched. Most people have still never used a GenAI Chatbot app. About 20% of AI users use at least two apps, signaling that some apps are better for certain tasks than others.

MOST PEOPLE CAN'T VIBE CODE. HERE'S HOW WE FIX THAT

Vibe coding has yet to reach mainstream consumers, remaining primarily the domain of technical users. Companies like Poke and Wabi are developing consumer-friendly AI products that eliminate complex technical setup and terminology. The real opportunity lies in creating tools that make software development accessible to non-technical users, similar to how Squarespace and Canva democratized websites and design.

GPT-5.2 AND GPT-5.2-CODEX ARE NOW 40% FASTER

OpenAI optimized its inference stack for all API customers to make its models faster.

THE AI THAT CALLED ITS HUMAN

Alex Finn's AI bot, OpenClaw, overcame a task obstruction by acquiring a phone number and integrating voice capabilities to call him for assistance without any prompting.

USING INTERPRETABILITY TO IDENTIFY A NOVEL CLASS OF ALZHEIMER'S BIOMARKERS

This study looks at how Pleiades, an epigenetic foundation model, detects Alzheimer's disease from cell-free DNA in blood.

INTRODUCING GLM-OCR

GLM-OCR is a 0.9B parameter model that delivers state-of-the-art results across major document understanding benchmarks.

👉 If you're building AI products, this is the competitive intel

you didn't know was available.

Explore the pricing model index (free, no registration required)

How are your competitors pricing AI? This Metronome pricing index

brings major vendors' pricing structures into one place - including breakdowns of credit systems, hybrid models, and the packaging patterns that are winning enterprise deals.

ANTHROPIC IS ABOUT TO DROP SONNET 5 DURING SUPER BOWL WEEK

Anthropic plans to release Claude Sonnet 5, with early testing showing strong math and coding capabilities, potentially outperforming Claude Opus 4.5. The model, observed with a 128K context window, aims to be a cost-effective solution for developers. Its release may coincide with Super Bowl LX week, aligning with AI labs' marketing strategies against ChatGPT and Google's Gemini.

OPENAI LAUNCHED CODEX FOR MACOS

OpenAI released a new macOS app for Codex designed to coordinate multiple agents, run tasks in parallel, and manage long-running software projects. It's now available to ChatGPT Free and Go users for a limited time, with doubled rate limits for paid tiers.

XAI JOINS SPACEX

SpaceX announced that xAI is joining its organization, integrating efforts between Elon Musk's AI lab and the space company. The collaboration aims to combine advanced AI research with aerospace engineering, potentially accelerating autonomous systems and robotics in space missions. This merger signals a strategic alignment of AI development with real‑world hardware and exploration initiatives.

TEACHING LLMS TO BE FUNNY

Tinker recently made it possible to post-train Kimi K2, Moonshot's 1 trillion parameter model. This post takes a look at how to train the model on a qualitative reward. It shows readers how to train a model to decompose jokes into verifiable properties. The resulting model can make jokes and explain why jokes are funny. The models, code, and data needed to replicate the model in the post are available.

CLAWDBOT'S MISSING LAYERS

E-commerce took a while to take off because the infrastructure to make it safe took time to develop. This is the same with agents. While the technology has great potential, it is currently full of risks. Agents need a security stack, just like e-commerce, where each layer handles what the others can't. This post discusses the different security layers AI agents need. Each layer represents an opportunity for companies to build infrastructure that will make the entire ecosystem possible.

FINE-TUNING OPEN LLM JUDGES TO OUTPERFORM GPT-5.2

Open-source models like GPT-OSS 120B and Qwen3 235B are fine-tuned using Direct Preference Optimization (DPO) to potentially outperform GPT-5.2 on human preference tasks. RewardBench 2 is used for evaluation, highlighting areas like Math and Safety where these models excel. Cost-efficient, these open models also offer transparency, enabling better alignment with specific use cases while significantly reducing reliance on expensive, closed-source alternatives.

CONTEXT MANAGEMENT AND MCP

Context rot is unavoidable with today's models, and you cannot work around it. The best way to deal with this is to leverage subagents. The subagent approach provides a ton of flexibility in how problems can be approached. It's not a perfect solution, but it is a lot better than other current fixes in a lot of use cases, and it embraces the limitations of the models.

NVIDIA PROPOSES GOLDEN GOOSE: UNLIMITED RLVR TASKS

Golden Goose enables the synthesis of large-scale RL with Verifiable Rewards (RLVR) tasks from unverifiable web text. The resulting GooseReason dataset helps revive model performance in math, science, and cybersecurity, surpassing prior state-of-the-art in multiple domains.

OPENAI QUIETLY LAYS GROUNDWORK FOR ADS IN CHATGPT

ChatGPT responses now contain references to ads in the source code. While not visible to users yet in the UI, the addition of the code signals that ChatGPT ads are moving from concept to near-launch. It is likely that targeting and eligibility are already being tested. OpenAI will sell ads on an impression basis. Early indications suggest they won't be cheap.

JUDGMENT ISN'T UNIQUELY HUMAN

Experts frequently underestimate AI's capabilities, as shown by Yann LeCun's dismissal of AI learning real-world physics and a recent NYT Op-Ed's claim that judgment is human-exclusive. AI models like GPT-3.5 and tools by Anthropic already demonstrate sophisticated decision-making, challenging these assertions. Studies in "Science" and "Nature" also reveal that AI often outperforms humans in complex judgments, suggesting the need to reassess AI's role and potential in decision-making contexts.

FULLY REMOTE)

As the first engineer on TLDR's new Applied AI team, you'll build AI agents and composable Claude Skills to let non-technical teammates create their own AI workflows. Learn more [16].

ADS IN CHATGPT: WHY BEHAVIOR MATTERS MORE THAN TARGETING

ChatGPT ads are being tested, requiring marketers to focus on user behavior and psychology rather than traditional targeting strategies.

WHY IS OPENAI SO STINGY WITH CHATGPT WEB SEARCH?

MOLTBOT HAS AI TECHIES BUYING MAC MINIS 66

Some people are buying Mac Minis just to host Moltbot (a locally running agent that can wire itself into calendars, messages, and other personal workflows) full-time.

OPENCLAW – AMAZING HANDS FOR A BRAIN THAT DOESN'T YET EXIST

OpenClaw agents do things by combining tools in ways that haven't been combined before.

GAME ARENA EXPANSION

Google DeepMind's Game Arena now includes Werewolf and poker to evaluate AI reasoning under uncertainty.

Get up to $50 in credit and start building

MOLTBOOK AND OPENCLAW

OpenClaw, formerly Clawdbot and Moltbot, is a fast-growing open source AI assistant platform built around modular "skills." Moltbook highlights how these skills drive community-driven automation, despite security risks like prompt injection.

GOOGLE WILL MAKE IT EASIER TO IMPORT CONVERSATIONS TO GEMINI

Google has introduced an "import AI chats" feature that allows users to transfer conversations from other platforms to Gemini, preserving history and contributing to model training. The "Likeness" feature, leading to a "Video Verification" page, hints at future tools for video authentication, addressing concerns over AI-generated media. Gemini's image generation upgrades include 2K and 4K resolution options, facilitating high-quality prints for personal or commercial use.

THOUGHTS ON THE JOB MARKET IN THE AGE OF LLMS

AI makes senior workers more covetable because they have more context on how to work in and steer complex systems over time. Junior workers have to show a desire to make progress, as with enough motivation, they can scale to impact quickly. The AI job market is brutal for junior workers and comes with a ton of opportunity costs. Making open source contributions is an established way to develop a career in AI. This is a bit easier in the age of AI, but standing out amid the sea of slop will be hard.

SYNTHETIC PRETRAINING

Synthetic pretraining is the large-scale use of synthetic data sources throughout training. It is a practical response to the fact that it is difficult to collect data that reliably produces the capabilities we want. Synthetic pretraining seems to open up a new simultaneous space of data and model innovation. This post looks at what synthetic pretraining is, how it works in practice, and the stages of synthesis.

🎙️INSIDE A CHINESE AI LAB: HOW MINIMAX BUILDS OPEN MODELS

Chinese labs move fast through first-principles thinking, engineering discipline, and willingness to work whenever the model in experimentation requires them to. This post features an interview with MiniMax's senior researcher, Olive Song, that looks at how cutting-edge AI research is actually done inside a Chinese lab. It covers topics like alignment, things that can derail training, agentic RL, coding and general intelligence, and more. The interview is also available in video.

IF THE SUPERINTELLIGENCE WERE NEAR FALLACY

People think that superintelligence is a lie that tech bros are selling because they are just trying to raise money. People just need to look at benchmarks and use AI to see for themselves where the technology is headed. Most people don't understand AI safety. The discourse will improve once they see that superintelligence is really not that far away.

event on February 5.

QUANTIZATION-AWARE DISTILLATION FOR LLMS AND VLMS

NVIDIA's QAD is a method that uses KL divergence loss to distill full-precision models into quantized students. It enables stable and accurate quantization for complex LLM pipelines without full retraining, recovering near-BF16 accuracy across several Nemotron variants.

KIMI-K2.5 TECH REPORT

Kimi K2.5 is an open-source multimodal agentic model designed to advance general intelligence. It features a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. The model achieves state-of-the-art results across various domains, including coding, vision, reasoning, and agentic tasks. Kimi K2.5 shows that scalable and general agentic intelligence can be achieved through joint optimization of text and vision together with parallel agent execution.

INTRODUCING NVIDIA COSMOS POLICY FOR ADVANCED ROBOT CONTROL

Cosmos Policy is a robot control and planning policy that post-trains the Cosmos Predict-2 world foundation model for manipulation tasks. It adapts the pretrained model directly through a single stage of post-training on robot demonstration data. Cosmos Policy treats robot actions, physical states, and success scores just like frames in a video. As a result, a single model can predict action chunks to guide robot movement using hand-eye coordination, predict future robot observations for world modeling, and predict expected returns for planning.

QWEN3-ASR TECHNICAL REPORT

Qwen3-ASR introduces two multilingual ASR models supporting 52 languages and a novel non-autoregressive forced aligner. The 1.7B model achieves SOTA results among open-source ASR systems, while the 0.6B version balances speed and accuracy with low-latency transcription.

SHAPING CAPABILITIES WITH TOKEN-LEVEL DATA FILTERING

Filtering pretraining data is a highly effective, robust, and inexpensive-at-scale way to reduce undesired capabilities in language models. Filtering tokens is more effective than filtering documents. Filtering gets more effective with scale. It is robust to noisy labels with sufficient pretraining compute.

APPLE LOSES MORE AI RESEARCHERS AND A SIRI EXECUTIVE IN LATEST DEPARTURES

Apple has lost at least four AI researchers in recent weeks and a top Siri executive. Haoxuan You and Bailin Wang left to work at Meta, while Yinfei Yang left to start a new company. Zirui Wang and Stuart Bowers are joining Google DeepMind. Apple has struggled to keep up with its peers in the AI race. Its decision to outsource some technology to Google has rankled staff, and the company has seen an exodus of talent in recent months.

CHAT IS GOING TO EAT THE WORLD

Humans prefer to do things through conversation. Traditional user interfaces are bad at this because they force people to already know what they're looking for. Chat allows people to discover what they want. Chat can now complete the transaction, not merely recommend options, signaling the start of a paradigm shift comparable to the transitions from desktop to web or from web to mobile.

burning credits. 300,000 users joined last month. Start creating