X / Twitter
Anthropic's Official Claude Account
Anthropic released an important feature for collaboration across models: on the Claude platform, a Sonnet or Haiku agent facing a complex decision can consult Opus through the advisor tool, obtain guidance, and continue executing, all within a single API request. On SWE-bench Multilingual, Sonnet with an Opus advisor scored 2.7 percentage points higher than Sonnet alone while reducing cost per task by 11.9%, improving both performance and cost. The feature is available in beta on the Claude platform.
https://x.com/claudeai/status/2036503582166393240
https://x.com/claudeai/status/2042308627478773808
https://x.com/claudeai/status/2042308629940855048
Anthropic Researcher Alex Albert
Albert explained the advisor tool's central design: let Sonnet “ask a friend for help” on difficult subtasks by dynamically calling Opus. This reduces tokens spent on repeated attempts while improving overall accuracy, striking a better balance between performance and cost.
https://x.com/alexalbert__/status/2042329150086922574
Claude Code Engineer Thariq
Thariq shared two valuable Claude Code tips. First, prompting will become a foundational skill with substantial impact, like writing or public speaking: it is the basis of communicating with agents, and he aims to help more people master it. Second, Monitor Tool is an underrated feature. For example, ask Claude to start a dev server and use MonitorTool to watch its error logs continuously, enabling genuine background monitoring.
https://x.com/trq212/status/2042318547519762678
https://x.com/trq212/status/2042335178388103559
Anthropic Claude Code Engineer Cat Wu
Wu announced that Claude Code has greatly simplified integration with AWS Bedrock and Google Vertex, substantially shortening setup and lowering the barrier to enterprise adoption.
https://x.com/_catwu/status/2042345489778331915
OpenAI CEO Sam Altman
Altman announced a new $100 ChatGPT Pro subscription tier in response to strong demand from many users. He also praised the Codex team, saying Codex has received a great deal of love and usage is growing rapidly.
https://x.com/sama/status/2042342572958630332
Andrej Karpathy
Karpathy observed a widening divide in perceptions of AI capability, largely due to differences in recency and access tier. Many people try older free tiers, while frontline builders use the latest paid APIs, producing vastly different experiences. He also shared an analysis of why the OpenClaw moment mattered: for the first time, many nontechnical users who previously knew only ChatGPT saw agents independently completing real-world tasks, prompting the realization that “AI really can do this.”
https://x.com/karpathy/status/2042334451611693415
https://x.com/karpathy/status/2042341482531864741
Google Labs VP Josh Woodward
Woodward announced that all features of Google Gemini's Lyria 3 music-generation model are now free for everyone. The Gemini app has generated more than 100 million songs in under 50 days since launch. Woodward said more creative ways to interact with the model are coming.
https://x.com/joshwoodward/status/2042318148851188151
https://x.com/joshwoodward/status/2042291365107941669
Vercel CEO Guillermo Rauch
Rauch highlighted shadcn's remarkable growth: `npx shadcn init` runs once every second, illustrating the scale of new software creation. He also outlined his view that “Agentic Infrastructure is the future of cloud computing”: (1) dedicated infrastructure for coding agents such as Claude Code, Codex, and Cursor; (2) deploying agents directly to Pages; and (3) agent-readable Logs and Analytics. Together these form Vercel's strategy for the agent era.
https://x.com/rauchg/status/2042294997278736552
https://x.com/rauchg/status/2042358253510963384
Zara Zhang
Zhang shared a counterintuitive position on AI writing: she rarely uses AI to write because she enjoys writing and has high standards. Editing AI text would take her more time, reducing efficiency. She also showed two AI-created HTML slide decks: an interactive English-learning deck in Duolingo's style, and one based on Andrej Karpathy's knowledge-management system.
https://x.com/zarazhangrui/status/2042379962339246105
https://x.com/zarazhangrui/status/2042349097701765564
https://x.com/zarazhangrui/status/2042355625179709573
Y Combinator CEO Garry Tan
Tan explained how GStack skills work: agents can infer from Markdown context when to trigger a skill, without explicit user invocation, automatically recognizing the need. He also praised Anjney Midha as “the real deal” and encouraged engineers to reject the idea that “good work takes time” and instead accelerate with AI tools.
https://x.com/garrytan/status/2042061979997831556
https://x.com/garrytan/status/2042109985346490483
OpenClaw Founder Peter Steinberger
Steinberger dramatically revealed that they had “found the developer of GitHub MCP Server.” The post received 2,675 likes and became a topic of the day. He also solicited questions about ClosedClaw and thanked contributor @thsottiaux.
https://x.com/steipete/status/2042214825405661677
https://x.com/steipete/status/2042184777575367117
OpenAI VP of Science Kevin Weil
Weil shared another AI mathematics breakthrough: AI solved five Erdos problems simultaneously. As models improve, their proofs are also becoming more elegant—getting the answer right and doing so beautifully.
https://x.com/kevinweil/status/2042073869880848481
Podcasts
Unsupervised Learning — Ep 84: OpenAI's Chief Scientist on Continual Learning Hype, RL Beyond Code, & Future Alignment Directions
OpenAI Chief Scientist Ako Paioki comprehensively discussed OpenAI's current research and views on AI's future in this unusually detailed interview offering an inside perspective.
Key takeaway: OpenAI's chief scientist believes AI is already intelligent enough to transform the economy. The priority now is creating practical value in real-world tasks, beyond abstract benchmark improvements.
On model progress and timelines, Paioki said OpenAI is rapidly approaching “research-level AI.” He distinguished an “AI research intern,” which needs fairly specific technical direction, from a “fully automated AI researcher,” which can work independently toward a broad goal. He expects the former's capabilities to be largely achievable this year; a fully automated researcher needs more time. The team already uses Codex extensively for real programming work, itself a strong signal of improving AI capability.
Paioki explained why mathematics and physics serve as research guideposts. Mathematical answers can be verified precisely, and difficulty scales indefinitely, from elementary problems to IMO Problem 6, providing a clear way to measure capability. More importantly, mathematical reasoning strongly correlates with reasoning in AI research; top mathematicians and theoretical physicists have consistently supplied some of OpenAI's best researchers. OpenAI now considers models intelligent enough to shift attention from math benchmarks toward creating value in real economic activity and applied science.
Paioki believes RL works beyond code and mathematics, but defining success for long-term tasks is challenging. Even a clearly specified mathematics or programming problem leaves open what to do on day one if it requires a year of work. The difficulty of long-term planning and evaluation substantially overlaps with the difficulty of soft-skill tasks that are hard to verify. This is RL's real frontier, and he has seen encouraging signs from scaling RL in these broader domains.
Paioki sees continual learning as important, but not neglected to the extent suggested in outside discussions. It is a natural extension of OpenAI's current research. On long-term alignment, he emphasized generalization: when models face distribution shifts in new situations, what default values will they fall back on? This fundamental question requires continued research.
https://www.youtube.com/@RedpointAI
No Priors — The Agentic Economy: How AI Agents Will Transform the Financial System with Circle Co-Founder and CEO Jeremy Allaire
Circle co-founder and CEO Jeremy Allaire, creator of the USDC stablecoin, explained how combining AI agents with programmable money will reshape finance.
Key takeaway: AI agents need payment capabilities to operate autonomously in the real world, and stablecoins provide their “natural monetary rails.”
Founded in 2013, Circle originally sought to create “a dollar protocol for the internet”: just as email is a communication protocol, USDC would be a protocol for transferring value. Drawing on Austrian economics and full-reserve money, Allaire views fractional-reserve banking as inherently fragile and stablecoins as a modern route to full-reserve money. USDC is fully backed by short-term U.S. Treasuries, averaging about 13 days to maturity, and cash. Through BlackRock, it has established daily transparency disclosures. Regulations worldwide, including the U.S. Genius Act, recognize this as a compliant stablecoin architecture.
Allaire described USDC handling everything from 25-cent on-chain gaming transactions to electronic trade settlements worth hundreds of millions of dollars under the same protocol, just as email treats content of different sizes alike. Stripe and Shopify have integrated USDC for merchant payments; Visa uses it to move funds within its network; and Ramp recently made it central to corporate treasury management, supporting supplier bills and payroll in USDC.
Allaire finds the intersection of AI agents and stablecoins particularly exciting. Agents acting autonomously in the real world inevitably need payments, including paying other agents for services. Stablecoins are their natural payment rails: internet-native money with public APIs, permissionless access, and 24/7 availability, perfectly matching how agents operate. He also emphasized blockchain auditability in the AI era: all inputs and outputs are publicly inspectable, providing a trustworthy way to verify agent behavior.
Allaire described blockchain as an operating system, analogous to mobile operating systems, web platforms, and AI foundation models. Its distinguishing features are tamper-resistant, fully auditable code and inherent guarantees of computation and transaction integrity. As regulations such as the Genius Act take effect worldwide, he expects stablecoins to become a geopolitical instrument for exporting digital U.S. dollars globally, as well as a safer alternative financial architecture.