← all issues · about Sift

Sift.

Week 2026-42 · Oct 12–18, 2026

10 stories · 17/20 feeds live · $0.18 run

Week at a glance

Models & Research 3Tooling 4Infra 1Policy 1Business 1

hi 9 · lo 5 · avg 6.6

Models & Research3 stories

9
#1

Google DeepMind launches Gemini 4 Argon, now generally available

Google DeepMind introduced Gemini 4 Argon, detailing its capabilities, a robotics-specific variant, and responsible-deployment practices, and made it generally available via the Gemini API and a $100/month plan. Independent benchmarks such as CursorBench and Vals AI reportedly corroborated the lab's performance claims.

Major frontier model release with robotics variant and independent benchmark confirmation; high reader interest.

Ddeepmind.google ×3Bblog.googleLlatent.spaceSsimonwillison.net
8
#2

Meta ships frontier Muse Wonder models and open-weight Muse Sparkneeds verification

Meta released its Muse Wonder and Muse Wonder Full models, reported as Claude 5.5-tier frontier systems, alongside the open-weight multimodal mixture-of-experts Muse Spark. Third-party analysis scored Muse Wonder Max as roughly equivalent to GPT-6.1 Sol.

Return of competitive US open weights plus a frontier-tier closed model; strongly matches model-release interest.

Iinterconnects.aiLlatent.spaceHhuggingface.coSsimonwillison.net
5
#10

Physical AI and interpretability research: Dyna-2, robotics, and polysemanticityneeds verification

Dyna Robotics detailed Dyna-2 world-action models aiming for hour-long autonomous physical agents, while separate work used agents to challenge polysemanticity and improve interpretability and alignment. These reflect progress in physical AI and mechanistic understanding of models.

Capabilities and interpretability research of interest, though partly secondhand claims.

Llatent.space ×2Ddeepmind.google

Tooling4 stories

8
#3

Anthropic rebuilds source control and code review for agent-scale commits

Anthropic described building a custom Git alternative after daily merged code changes rose from about 2,500 in May 2026 to over 30,000 by September, and detailed continuous Claude-driven code review replacing end-of-pipeline human gates. Related reporting adds insider context on the migration and the rise of AI-assisted coding workflows inside big tech.

Concrete numbers on agentic SDLC at scale (2,500 to 30,000 merges/day); directly on-profile.

Ssimonwillison.net ×2Nnewsletter.pragmaticengineer.com
6
#6

Inside Uber's AI infrastructure and developer-tool workflowsneeds verification

A Pragmatic Engineer interview with Uber Distinguished Engineer Uday Kiran Medisetty details how Uber builds internal developer tools and integrates AI agents into engineering workflows. It also covers how the company measures agents' impact on productivity.

Concrete enterprise view of agent integration and productivity measurement in a major engineering org.

Nnewsletter.pragmaticengineer.com
6
#7

Frontier labs detail agent autonomy and workflow coordination at AI Engineer NYCneeds verification

At AI Engineer New York, OpenAI and Anthropic discussed giving agents useful autonomy while keeping consequential financial decisions under human control. Anthropic also drew on Unix and Linux lessons to explain dynamic workflows that coordinate Claude agents on larger projects.

Substantive agent-harness and autonomy design content from OpenAI and Anthropic, on-profile.

Llatent.space ×2
5
#8

Coding-agent guardrails: commit hooks, quizzes, and Codex crash reports

Developers shared approaches for controlling AI-generated code, including Sphinx, a hook that quizzes you on your agent's code before shipping, and pre-commit hooks that enforce library usage by failing commits and re-prompting the agent. Separately, a widely-discussed GitHub issue reports Codex Desktop continuously generating Crashpad dumps.

Practical agent-harness and SDLC-automation tooling discussion matching dev-tooling interest.

Ssphinx-six-theta.vercel.appSsimonwillison.netGgithub.com

Infra1 story

7
#5

Modal reaches $1B ARR and $16B Series D as AI infra decacornneeds verification

Modal reportedly reached $1B ARR and raised a $16B Series D, tripling usage and valuation in five months and growing 15x in a year. The report positions Modal as a leading sandbox-based AI infrastructure provider.

Significant AI compute/sandbox infrastructure milestone with concrete growth numbers.

Llatent.space

Policy1 story

5
#9

Google DeepMind and Anthropic deepen Singapore government AI work

Anthropic opened a Singapore office and signed an MOU with the Government of Singapore, while Google DeepMind and GovTech Singapore deployed fine-tuned on-device Gemma models to intercept conversational scams with zero cloud data egress. Both reflect expanding public-sector AI collaboration in the region.

Concrete government partnerships and an on-device model deployment, a secondary interest area.

Aanthropic.comDdeepmind.google

Business1 story

7
#4

Anthropic tells investors Q3 revenue hit $16.8B ahead of IPOneeds verification

Per an FT report, Anthropic told investors quarterly revenue rose nearly 50% to $16.8bn in Q3, up from about $4.8bn in Q1 and $11.5bn in Q2. The company also reported $1.70 in revenue for every $1 spent on computing power in Q3.

Concrete enterprise AI revenue and unit-economics figures, though sourced secondhand via FT.

Ssimonwillison.net

Sources scanned · 17/20 live