Everything that mattered in AI, one page a week.

Evo 2 is a genomic foundation model trained on the DNA of more than 100,000 species spanning all domains of life, on roughly 9 trillion nucleotides.

At 40B parameters it is described as the largest AI model for biology to date, with context up to 1 million nucleotides and a new StripedHyena 2 architecture. It can flag disease-causing mutations and generate coding-rich sequences up to a million bases long; Arc says 16 of 285 designed phage genomes were viable in lab testing. Training data, training and inference code, and model weights were all released, with the model integrated into NVIDIA BioNeMo alongside an interpretability visualizer.

Microsoft announced Majorana 1, a quantum processor it says is the first powered by a topological core, with eight topological qubits on a chip designed to scale toward a million.

The claim rests on a peer-reviewed Nature paper describing a hardware-protected topological qubit built from an indium-arsenide and aluminum stack Microsoft calls a topoconductor, plus a measurement technique precise enough to read the parity state of a superconducting nanowire. DARPA moved Microsoft into the final phase of its program to build a fault-tolerant prototype in years rather than decades. Microsoft says the chip will not be offered through Azure.

xAI unveiled Grok 3 and began rolling it out to X Premium+ subscribers, plus a new $30 per month SuperGrok tier.

xAI says the family was trained on its Colossus supercluster with roughly 10x the compute of previous state-of-the-art models — reporting put the Memphis cluster at around 200,000 GPUs — and that reasoning was refined through reinforcement learning at large scale, letting Grok 3 think for seconds to minutes. xAI claims an Elo of 1402 on Chatbot Arena and leads over GPT-4o on AIME and GPQA; Grok 3 mini targets cost-efficient reasoning, and DeepSearch shipped as a research agent. Grok 3 is closed, but Musk said Grok 2 will be open-sourced once Grok 3 is mature and stable, which he estimated at a few months. Two days later xAI made Grok 3 free to all users until our servers melt.

Perplexity released R1-1776, a post-trained version of DeepSeek-R1 whose weights can be downloaded and self-hosted.

R1-1776 keeps the base architecture: a 671-billion-parameter mixture-of-experts model with 37 billion active parameters and a 128K context window, released under the MIT license and also served through Perplexity's hosted API. Perplexity said it employed human experts to identify roughly 300 censored topics, built a dataset of 40,000 multilingual prompts, and fine-tuned using an adapted NVIDIA NeMo pipeline, reporting that math and reasoning performance held at parity with base R1.

OpenAI published SWE-Lancer, a benchmark of 1,488 freelance software engineering tasks drawn from Upwork and built around Expensify's codebase, representing $1,000,000 in real payouts actually made to the freelancers who did the work.

Tasks split into independent-contribution tickets — from $50 bug fixes to $32,000 feature implementations — and managerial decisions where the model picks between competing implementation proposals. Grading uses end-to-end tests written by professional engineers rather than unit tests. Claude 3.5 Sonnet did best, completing 26.2% of the hands-on coding tasks and 44.9% of the management decisions; OpenAI's own note was that frontier models are still unable to solve the majority of tasks.

At the Munich Security Conference, UK Technology Secretary Peter Kyle recast the AI Safety Institute as the AI Security Institute, days after the Paris summit.

The remit narrows to national security and crime, including chemical and biological weapons, cyber-attacks, fraud and child sexual abuse, with a new criminal-misuse team run jointly with the Home Office and a partnership with DSTL to assess frontier-AI risk. The same announcement carried a new agreement with Anthropic, published as a Memorandum of Understanding, to apply Anthropic's models to public services.

OpenAI published a major Model Spec update — 63 pages, up from roughly 10 previously — and released it into the public domain under CC0 so anyone can reuse or fork it.

The rewrite foregrounds customizability, transparency and intellectual freedom, meaning fewer arbitrary content restrictions while keeping guardrails against real harm, and it explicitly loosens how the model handles controversial topics and adult content. OpenAI also published early adherence results against the spec's principles. On the same day, Sam Altman laid out a product roadmap: GPT-4.5 in weeks, GPT-5 in months as a unified system collapsing the o-series into the mainline model.

OpenThoughts released OpenThinker-32B under Apache 2.0, built on a Qwen2.5-32B base and trained on OpenThoughts-114k — a dataset whose problems, ground-truth solutions, test cases and DeepSeek-R1 reasoning traces are all published, along with the data-curation code.

The team's point is methodological: scale the data, verify reasoning traces, scale the model, and an open-data 32B model outperforms existing open-data reasoning models on math, code and science — while the comparable R1-Distill-32B is a closed-data model. The same day, the Barcelona Supercomputing Center published Salamandra: 2B, 7B and 40B decoder-only models trained from scratch on 35 European languages plus code, released under Apache 2.0 with training and evaluation scripts public.

The AI Action Summit in Paris ended with more than 60 countries, the African Union Commission and the EU signing a declaration on inclusive and sustainable AI — and the United States and the United Kingdom declining to sign it, on grounds of over-regulation and national-security framing respectively.

The investment numbers were the concrete output: France announced €109 billion in private AI commitments, and the European Commission launched InvestAI, a €200 billion mobilization including a €20 billion European fund for AI gigafactories of roughly 100,000 chips each — about four times the size of the AI factories already being built.

Judge Stephanos Bibas of the US District Court for the District of Delaware granted Thomson Reuters partial summary judgment, holding that Ross Intelligence's copying of Westlaw headnotes to train a legal-research AI was not fair use.

The court rejected the transformative-use argument because the material was used to build a direct market substitute, and split the fair-use factors against Ross. It is the first US decision squarely addressing whether training an AI model on copyrighted material qualifies as fair use, and it landed while the larger training-data cases against OpenAI, Anthropic and others are still running.