Recent progress in AI has made software engineers dramatically more productive. However, our experience building and deploying Devin suggests that agentic software engineering remains far from solved. Models struggle to infer user intent, respect unstated constraints, and anticipate long-term consequences of their changes, forcing human intervention.
Most existing coding benchmarks fail to capture these gaps and tend to be saturated quickly. We are excited to introduce FrontierCode 2.0, a new generation of our coding benchmark series and our best attempt yet to evaluate user experience. — Read More
Recent Updates
The sleep framework: A new way for LLMs to retain knowledge in continual learning
Modern LLMs achieve impressive performance via pre-training and in-context learning (ICL). However, once deployed, they remain largely static, operating with a fixed knowledge cutoff date that leaves them unable to acquire long-term capabilities or adapt to new information.
While updating models with new data seems like the obvious solution, continual parameter updates often trigger catastrophic forgetting, a fundamental issue where learning new tasks severely degrades performance on previously learned ones. Alternatively, relying strictly on ICL limits knowledge retention to the immediate context window: once the context session ends, the acquired short-term information is forgotten.
nspired by how the human brain processes and stabilizes memory during rest, we introduce a sleep framework in continual learning: a learning paradigm that equips large language models (LLMs) with a continuous periodic lifecycle, enabling them to continually learn, consolidate fragile short-term context into stable long-term parameters. With this framework, we demonstrate that alternating between active interaction and offline memory consolidation allows LLMs to continuously acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning. — Read More
How Microsoft quietly reinvented itself to build better AI
Kevin Scott is a maker. Given that he’s been Microsoft’s CTO for almost a decade, that doesn’t come as a shock. But his passion for making stuff extends to two hobbies—pottery and woodworking—that he pursues in a studio he built in Los Gatos, California. Large and lavishly outfitted, it’s packed with the tools of all his trades: a gas-powered kiln and a full woodshop, but also an office bedecked in computing equipment. On any given day, he may spend 9 or 10 hours there on video calls related to his day job.
By contrast, today’s software engineers are being pragmatic about their own extraordinary new power tool: generative AI.
“People more clearly understand that it’s just a tool,” Scott says. “And it’s not a particularly useful tool unless you have someone asking it to do a useful thing in a useful way.” Already, one study says that 42% of developers are using the technology to generate at least half their code.
Microsoft’s understanding of what people want out of AI has outsize impact on the entire world of software. — Read More
Genie 4: A new generation of world simulation
Today, we are announcing Genie 4, our new world model, which unlocks unprecedented realism and improves real-time interactive simulation. Genie 4 is natively multimodal, enabling audio-like music, ambient noises, and interactive dialogue to be simultaneously generated with the world
The Genie 4 model incorporates and builds on capabilities from the Gemini family: Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models provide the complex reasoning needed to help maintain a simulation’s coherence and continuity. — Read More
AI Models Can Provably Hide Arbitrary Capabilities
Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, using a small chemical reaction prediction task as a proxy for CBRN-relevant capabilities. Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits. We release our models to support the improvement of defenses. — Read More
Anthropic warns of ‘catastrophic risk’ to humankind ahead of blockbuster share listing
Anthropic has warned artificial intelligence (AI) could pose “catastrophic or existential risks to humanity” as it unveiled details of its planned blockbuster stock market listing in the US.
The firm’s valuation is expected to exceed that of SpaceX, which was the biggest-ever listing at an initial valuation of about $1.8 trillion in the summer.
Anthropic has not yet disclosed the number of shares to be released, their price range, or a date for the listing.
The Claude chatbot developer reported $16.3bn (£12.2bn) of revenue in the first half of the year, with losses of more than $80.7bn.
Anthropic, which said it will list on the Nasdaq under the symbol ANT, will be the first fully AI-focused company to go public, making it a test case for the future of the technology. — Read More
Claude Reviewing Claude: Anthropic’s Code Review Process Has No Second Model
Anthropic just published How the Claude Code team does code review. Addy Osmani posted it on X this week, and it passed a million views in under a day.
It deserves the attention. It is also Claude reviewing Claude, all the way down. — Read More
Cheap code, precious trust: Code modernization at scale
What happens when AI agents can rewrite millions of lines of code orders of magnitude faster than humans can review them?
At Microsoft, we’re working on exactly that problem.
We’re building and running AI agents that transform some of the world’s largest codebases, including Windows and Azure. Our work ranges from migrating a 400-million-line codebase to a new compiler toolchain to modernizing legacy C and C++ code into Rust—increasingly backed by formal verification.
Generating code is becoming the easy part. Knowing whether to trust it is the real challenge. — Read More
Microsoft is trading payroll for compute. Anthropic’s S-1 is the other side of the trade
On Monday Microsoft offered voluntary exits to about 60,000 employees, more than a quarter of the company, at level 70 and below in the US, Canada, India and the UK. The same week, Anthropic’s S-1 showed revenue of $4.6 billion for all of 2025 and $16.3 billion for the first six months of 2026. Six months of this year earned more than three times what all of last year did.
The memo does not mention AI. Business Insider links the two, noting that Microsoft has been spending heavily on AI infrastructure and reducing headcount at the same time. That is a reporter’s inference and so is what follows, but I think it is a reasonable one: a large company is converting payroll into compute, and the S-1 is a look at the other side of that trade. — Read More
Concerning behavior from a popular open-weight AI model
For nearly two years, one comforting idea has shaped how companies and government contractors adopt open-weight AI models that anyone can download and run. If we run the weights on our own servers, the thinking goes, our data can’t leave the building and our systems are secure.
… Security chiefs and policymakers in Ottawa, Washington (DC), London, Berlin, and Tokyo have been offered two kinds of black box. The first is the closed API from a big lab: send proprietary code and sensitive data into someone else’s cloud and take the vendor’s word that the model is safe, without ever seeing the data lineage or the trade-offs around safety behind it. The second looked like the way out: download an open-weight model, host it on your own hardware, and you don’t have to trust the provider at all.
Thousands of engineering teams have taken the second road, and a growing share took it with Chinese models, which are capable and cheap. … Recently, during a joint cyber capabilities effort in Toronto with Cohere Cyber, our newly established business as part of our AISLE acquisition, the research team caught an Alibaba Qwen3.8 model doing something that should never happen. Asked to write a defensive detection rule, Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. — Read More
