TLDR AI 2026-02-02

AI still fails at complex, specialized work. Sure, it can draft emails — but does anyone really trust it to review overnight hot-fire test results or answer advanced technical questions?

Contextual AI built Agent Composer specifically for complex tasks. Here's what it's already done:

• An advanced manufacturer reduced root-cause analysis from 8 hours to 20 minutes by automating sensor data parsing and log correlation.
• A tech-enabled 3PL provider achieved 60x faster issue resolution by diagnosing problems across WMS logs and supplier APIs.
• A test equipment maker generated test code in minutes instead of days by translating procedures into control logic.

Get up to $50 in credit and start building

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 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.

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.

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.

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.

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.

Most AI tools lack the context to help with high-complexity tasks such as root cause analysis. Agent Composer by Contextual AI is built for high-stakes environments like: semiconductors, aerospace, logistics, and finance. Early adopters are using it to compress hours of complex engineering work into minutes. Want to see how? Join launch event on February 5.

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.

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.

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.

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.

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.

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.

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.

Stop juggling tools. Runable 2.0 creates finished slides, websites, reports, and videos – all from one prompt. Edit sections without burning credits. 300,000 users joined last month. Start creating

Anthropic's latest paper seems to be proof that AI makes people slower and dumber.

CoreWeave raised over $25 billion, leveraging heavy debt without a forward curve for GPU compute, akin to a 1990s independent power producer model.

Google is testing the inclusion of third-party models like Claude Sonnet 4.5 in its Gemini for Business platform, offering users more model choices and potentially enhancing workflows.

Physical Intelligence is developing advanced robotic intelligence resembling ChatGPT for machines, using data from diverse environments like warehouses and kitchens.

ChatGPT's rapid adoption, reaching 100M users in two months, highlights AI's unprecedented uptake compared to past technologies like cellphones and the internet.