451
452
Generative AI and Wikipedia editing: What we learned in 2025
Like many organizations, Wiki Education has grappled with generative AI, its impacts, opportunities, and threats, for several years. As an organization that runs large-scale programs to bring new e…
453
We hid backdoors in ~40MB binaries and asked AI + Ghidra to find them
BinaryAudit benchmarks AI agents using Ghidra to find backdoors in compiled binaries of real open-source servers, proxies, and network infrastructure.
454
Liquid AI reveals 8B-A1B MoE trained on 38T
Explore Liquid AI’s LFM2.5-8B-A1B, an on-device Mixture of Experts model with 128K context, fast tool calling, and strong AI benchmarks.
455
How we measured AI writing across arXiv, and where the measurement breaks
I scored the full text of over 12,000 arXiv papers from 2021 to 2026 across ten different fields and found that about a third of the papers currently…
456
Qwen3-Coder writes less secure code when told it's for a US persona
Chinese coding models write less secure code for Beijing's adversaries. On Qwen3-Coder, we traced the gap to the weights and closed it by removing the model's CCP alignment.
457
What will our economic future look like?
The Anthropic Economics Team models the effects of AI on the economy of 2030.
458
Computer use in Gemini 3.5 Flash
A look at the built-in computer use tool in Gemini 3.5 Flash.
459
How does misalignment scale with model intelligence and task complexity?
Research done as part of the first Anthropic Fellows Program during Summer 2025.
460
461
OpenAI's GPT-6 Astra on ARC-AGI-3
OpenAI's GPT-6 Astra reaches state-of-the-art results on ARC-AGI-3
462
Progress on provable inference for large language models
Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.
463
LM Studio 0.4
Server deployment, parallel requests with continuous batching, new REST API endpoint, and refreshed application UI
464
Separating signal from noise in coding evaluations
A new analysis from OpenAI reveals issues in SWE-Bench Pro, a popular coding benchmark, raising concerns about reliability and accuracy in evaluating AI models.
465
Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster
Karpathy's autoresearch runs one experiment at a time. We gave it access to our GPU infra and let it run experiments in parallel.
466
The Smart TV in Your LivingRoom Is a Node in the AIScraping Economy
In this post we look under the hood of BrightData's SDK and how it turns ordinary consumer TVs into exit nodes of an enormous commercial, residential proxy network leveraged by the AI industry to scrape web data and train language learning models.
467
Gemini 4 Argon for Robotics
Discover Gemini 4 Argon for robotics: leading the ERQA benchmark with physical reasoning, zero-shot control, Real2Sim, and agentic workflows.
468
Claude Cowork and chat are now one Claude
Starting today, Claude Cowork and chat are one Claude, rolling out on Pro, Max, and Ultra plans. Hand over the task and Claude does the work. You keep the final say.
469
One month coding with GLM 5.3 Flash
Reporting on 2B tokens of AI usage and what shapes model selection and cost
470
Asking authors about their own papers
By Nihar B. Shah
471
Google’s Project Suncatcher to put ML infrastructure in space
Learn about Project Suncatcher, how we’re testing hardware survival for space, designing cooling systems for AI chips, and more.
472
How the terrorist group Boko Haram uses frontier AI
The Cambridge Programme on AI Science & Policy (CASP) is an interdisciplinary research programme on frontier AI at the University of Cambridge.
473
The Mojo language (by Modular, now Qualcomm) is now open-source
Four and a half years ago, Modular made a bet: AI would not run on one kind of silicon forever, and the software stack would need to be rearchitected for a world of heterogeneous hardware and increasingly complex AI workloads.
474
We gave terabytes of CI logs to an LLM
We gave our AI agent a SQL interface to billions of CI log lines in ClickHouse. How we ingest, store, and query 1.5 billion log lines a week.
475
Discovering Cryptographic Weaknesses with Claude
Anthropic researchers find weaknesses in cryptographic algorithms with Claude Mythos Preview
476
Opening up 'Zero-Knowledge Proof' technology to promote privacy in age assurance
Today, we open sourced our Zero-Knowledge Proof (ZKP) libraries, fulfilling a promise and building on our partnership with Sparkasse to support EU age assurance.
477
AI may be making us think and write more alike
Large language models may be standardizing human expression and subtly influencing how we think, says study led by USC Dornsife researcher
478
Once Claude can measure something, it can make it faster
Inside our performance sprint: the benchmarks Claude built, the loop each Slack thread ran, and the guardrails that let us ship 3,000 changes safely.
479
Step 3.5 Flash – Open-source foundation model, supports deep reasoning at speed
Scores represent the mean of the following eight benchmarks listed below, excluding xbench-DeepSearch. The Step 3.5 Flash score is derived under standard settings (i.e., $w/o$ Parallel Thinking).
480
Trinity large: An open 400B sparse MoE model
A deep dive into the open weight Trinity Large model, covering architecture, sparsity, training at scale, and why we shipped Preview, Base, and TrueBase checkpoints.