Parse notes
Sponsored
Trusted by OpenAI, NVIDIA, and Anthropic, Metronome is the infrastructure for modern monetization. As part of a week's worth of launches, they're announcing a host of new capabilities - including three new ways to support flexible and hybrid pricing models [4]:
Google plans to launch Gemini 3 and Nano Banana Pro next week. The company has consistently updated its Gemini models across its suite simultaneously, and the same broad rollout is expected again this time. The Pro branding suggests that Google plans to provide accessible and production-grade generative tools across its platform ecosystem.
GPT-5.1 is now rolling out to users. It features 8 unique chat styles, making it easier for users to set the tone and style that feels right for them. This Reddit Ask Me Anything features members from OpenAI's team.
FutureHouse (which is transitioning to commercial spinout Edison Scientific) released an AI scientist that reads 1,500 papers and runs 42,000 lines of analysis code per run. Beta users estimate that a single Kosmos run accomplishes what would take them six months, with 79.4% of conclusions accurate, though the system sometimes pursues statistically significant but scientifically irrelevant findings. It's already made seven discoveries—three reproduced unpublished findings and four contributed novel insights, including one related to Alzheimer's that was validated in human tissue. Kosmos is launching at $200 per run with a generous free tier for academics.
Most modern language models require massive computational resources and months of training time. This post discusses how to achieve the optimal recipe for creating efficient pre-training data sets to achieve high performance while using dramatically less data. It uncovers critical insights about dataset mixing strategy, failure modes in curriculum testing, and the 'goldilocks zone' for synthetic content. The resulting mixture consistently outperforms complex curriculum strategies and avoids catastrophic failures while maintaining excellent generalization.
A race is a competition with clear boundaries and a clearly defined finish line. This isn't the case with the AI 'race' between the US and China. Nobody currently knows where AI is headed. The field is still being explored. The US and China are taking different strategies that play to their unique strengths. They are locked in a structural conflict, and harmony appears impossible.
Sponsored
Most teams ask "Which tool is best?" The better question: "Which approach gets us from idea to collaboration fastest?" Miro's AI prototyping landscape guide [10] provides a requirement question set, decision framework, and evaluation scorecard so you can pick the tools
Structured outputs are now available in public beta for Claude Sonnet 4.5 and Opus 4.1. Support for Haiku 4.5 is coming soon. Without structured outputs, Claude may generate malformed JSON responses or invalidate tool inputs, breaking applications, even with careful prompting. Structured outputs constrain Claude's responses so they follow a specific schema. This ensures valid, parsable output for downstream processing.
The context window is the entire input that a large language model receives when generating output. It contains messages, the model's replies, tool calls, and the thinking blocks the model outputs to 'reason'. Everything that happens between the user and the model goes into the context window. This article walks users through some of the ways Amp gives users to manage the contents in a context window.
Depth Anything 3 (DA3) estimates spatially consistent depth from single or multi-view images, regardless of camera pose availability. It relies on a plain transformer backbone and a single depth-ray representation to simplify architecture and training.
GAD introduces a new approach for distilling knowledge from black-box LLMs using only generated text, without access to output probabilities. It trains a student model through adversarial learning against a discriminator, enabling on-policy imitation.
AI allows us to write programs by specifying objectives. If a task is verifiable, then a neural net can be trained to do it extremely well. The more a task is verifiable, the more it is amenable to automation using AI.
Microsoft co-designed its new AI superfactory with OpenAI and Nvidia, integrating three generations of frontier training systems and adding AMD for GPT inference to broaden the hardware stack. The company aims to scale this infrastructure so enterprises can train and run their own models without ceding value or control to major AI vendors. The biggest impact will come when industries like pharma, manufacturing, and education use AI to compress timelines and expand output, creating positive-sum economic gains.
Sponsored
Cisco's AI Networking webinar [17] reveals how to build lossless, scalable Ethernet fabrics for optimal GPU utilization, lower latency, and simpler operations —Register now [17] and power up your AI data center!
Yann LeCun may soon be leaving Meta to pursue a startup focused on world models because he believes that large language models are a dead end in the pursuit of superintelligence.
Sakana AI raised $132 million from investors including MUFG, Santander, and US venture firms, pushing its valuation to approximately $2.6 billion, which is the record for an unlisted Japanese startup.
Google Maps introduced AI tools, including a builder agent and an MCP server, enabling developers to create interactive projects using Maps data.