From the archive
AI research tests new methods for training, evaluating and understanding artificial intelligence. Find news on machine learning papers, reasoning, model behaviour and experiments that challenge existing assumptions. Read what researchers measured, how they tested it and where the limitations lie, with links to the original work and discussion of what the findings could mean beyond a benchmark.
Lead story
Tata Elxsi says its IRIS platform uses edge computing and computer vision to spot industrial safety risks in seconds, rather than leaving operators to discover them minutes later. The important change is practical: cameras become an active warning system, while most raw video stays on site.
Read the report & discussion ↗DeepSeek Watch
DeepSeek-V4.1-Flash is a 552-billion-parameter mixture-of-experts model built for long-context, tool-using agents, with a one-million-token context window and design choices aimed at reducing memory pressure. The important point is not that the model is small. It is that DeepSeek is trying to make a very large model more practical…
Watch Desk
Jared Palmer has released Kev, a family of small AI decision models that return answers such as yes/no probabilities, multiple-choice decisions and ratings instead of conversational prose. The practical point is compellingly unglamorous: developers can run these models locally for fast, repeated decisions without sending every request to…
Watch Desk
Snorkel AI has raised $350m at a $3.5bn valuation, according to Reuters, putting a substantial price tag on the less glamorous work that makes AI systems useful: creating and checking the data they learn from. The company says its platform uses people and AI agents to develop and…
OpenAI Watch
Researchers at cybersecurity company Hacktron used Anthropic’s Claude to help exploit vulnerabilities affecting OpenAI’s community forum and an OpenAI single sign-on system, eventually reaching an internal software repository. The incident matters because AI did not replace the security researchers, but it appears to have made a difficult exploit…
Watch Desk
OpenRouter has launched a Batch API that cuts inference prices by half for users willing to let requests run asynchronously within a 24-hour window. The service is aimed at large, non-urgent workloads such as bulk text generation and embeddings, where immediate responses are less valuable than a lower…
Mistral AI Watch
Mistral has patched a prompt-injection technique in its Le Chat assistant that could have exposed users’ personal data to an attacker-controlled server. The company says exploitation required a person to paste an adversarial prompt and provide personal information, and that no users were affected. That makes this a…
Watch Desk
Machine-learning systems can quietly become less reliable even when their code, weights and confidence scores stay exactly the same. A new explainer by Cristobal Santana sets out how changing data distributions can cause that silent failure, and why checking the data matters as much as checking the model.
Watch Desk
Yandex has released the weights for Alice AI Foundation, an 80bn-parameter language model that the company says is available for commercial and research use. It is a substantial open-model launch, but also a revealing bid to build a more self-sufficient Russian AI stack.
Alibaba Qwen Watch
An exploratory study suggests Qwen models may contain internal representations of older language models, including GPT-2. The intriguing finding is not that Qwen can imitate GPT-2, but that its completions may track GPT-2’s hidden continuation more closely than Qwen’s own usual output.
Watch Desk
StepFun has released Step 5 Preview, a large mixture-of-experts model offering a one-million-token context window, image input and API pricing of $1 per million input tokens and $2.70 per million output tokens. The practical significance is not merely that another frontier model has arrived. It is that long-context,…
Watch Desk
A new tool called Product Prover applies formal-verification ideas to ordinary product specifications, using an LLM to reconstruct state machines and flag missing rules before developers generate code. That makes it a useful AI engineering story because it targets one of software’s oldest problems: building precisely what nobody…
Watch Desk
South Korean semiconductor-equipment maker Melcon plans a KOSDAQ listing on 15 October, aiming to raise up to 30.8 billion won for High-NA EUV development, extra production capacity and expansion into TSMC and HBM-related work. It is a useful reminder that the AI hardware race also depends on the…
