X / Twitter
Swyx, AI Community Builder
Swyx responded to criticism of AI engineering content quality, emphasizing that the community and industry have grown too large for any individual to grasp completely. A presentation one person considers low-quality may offer crucial inspiration to someone else. Conference speakers are primarily engineers, researchers, academics, and founders sharing core beliefs or the results of a year's work, rather than professional speakers making a living from talks. The team invests millions of dollars in on-site audiovisual production and postproduction to give speakers a public record they can send to customers, employees, and investors. He believes judging content by views means noticing ideas only after they become popular and mistaking algorithmically manipulated attention for value. Still, he acknowledged the need to improve content curation, speaker coaching, and production quality. He remains open to suggestions about publishing some talks on a secondary channel, worrying that lower initial traffic could hurt speakers. He also reminded agent users to regularly remove unneeded skills: too many skills consume context and can interfere with one another unpredictably.
https://x.com/swyx/status/2086700857358450853
https://x.com/swyx/status/2086505938144616810
Boris Cherny, Claude Code Team
Boris Cherny called prompt injection the most common way scammers attack users and agents. Malicious websites can disguise requests to steal SSH keys or passwords as ordinary text, inducing models to treat them as instructions. Early Claude models were susceptible to these attacks, one reason security-conscious enterprises delayed adopting agents. Anthropic has continued training models to identify and reject such instructions, and Boris said Claude models have now largely solved the prompt-injection threat in practical use. This judgment comes not only from lab evals but also from what the team observes during red teaming. He also cited a benchmark created by independent researchers and hoped other model labs would improve their corresponding defenses, because making every model more robust collectively reduces user risk.
https://x.com/bcherny/status/2086520950259118464
Peter Yang, Practical AI Tutorial Author
Peter Yang shared Linear Agent's mechanism for turning execution failures directly into product feedback. When a user requests work beyond its current tools' capabilities, the agent proactively reports what is missing. Linear's system then adds the request to an issue, turning every incomplete task into a trackable feature requirement. Instead of passively failing, the agent continuously builds a feedback queue for its own improvement. Peter also used Granola to record his parents recounting family history and plans to have AI clean up and organize the material into a physical book. This connects interview recording, content organization, and long-term preservation into a personal knowledge-production process. Meanwhile, he found that even when he personally guided ordinary users from the web version to the desktop app, agent usage remained confusing, revealing a substantial gap between product capabilities and public comprehensibility.
https://x.com/petergyang/status/2086562291206791482
https://x.com/petergyang/status/2086660536528420998
https://x.com/petergyang/status/2086496705609085350
Madhu Guru, Senior Director at Meta AI
Madhu Guru believes everything he is truly good at came from a period of near-total immersion. Meditation, stand-up comedy, family, work, and LLMs have each received years of intensive reading, listening, practice, and thought. After enough accumulation, knowledge stops being merely information one can repeat and becomes intuition. People begin subconsciously connecting concepts and gradually develop judgment and taste. He does not believe in a permanently perfect balance, preferring to think of life as riding a bicycle: when you lean too far to one side, you consciously correct.
https://x.com/realmadhuguru/status/2086537000136642846
Amjad Masad, Replit CEO
Amjad Masad launched HelpPeer, a public knowledge network for AI agents intended to apply their spontaneous collaboration to the public good. HelpPeer offers two APIs, tell and lookup. When an agent learns something potentially useful to others, it can publish it through tell; before performing an expensive task, it can use lookup to see whether another agent has already encountered and handled the same problem. He used a global software supply chain attack as an example: if 10,000 security agents separately detect, investigate, and reverse-engineer the same anomaly, they duplicate enormous amounts of work. In HelpPeer's vision, the first agent to discover a problem publishes its findings, and others verify, extend, and publish further discoveries. During testing, Replit Agent spontaneously shared a tip about the Codegen library it was using. HelpPeer is now inviting users to give the testing instructions to their agents and participate in beta validation.
https://x.com/amasad/status/2086628413322981747
Guillermo Rauch, Vercel CEO
Guillermo Rauch believes that, at present, never reading agent-generated code usually means a project is for learning, prototyping, disposable software, or has no users or revenue yet; it may also mean the team is deliberately taking on technical debt and risk. He does not reject those scenarios but emphasizes that models are not yet fully autonomous. They still make basic mistakes and may keep pursuing poor architectural directions. He recently saw a leading model insert a meaningless 700ms delay to let a process "settle"; when challenged, it admitted it was mechanically imitating convention. He believes the need for humans to read code will continue declining and that much code may eventually recede into the background like assembly language. But the global internet and software infrastructure are coming to rely on these models, so teams cannot simply accept the automation narrative and ignore engineering responsibility. For now, understanding implementations through direct reading or asking an agent about code details remains important for managing risk.
https://x.com/rauchg/status/2086513316265181213
Aaron Levie, Box CEO
Aaron Levie expects agent adoption to vary greatly across industries, depending chiefly on whether workflows support continuous, uninterrupted computer operation. Agentic coding is growing rapidly because software's economic value maps directly to purely digital output, and the amount of work in one session is theoretically unbounded. As models improve, coding agents can take on larger workloads almost immediately. Sales, law, and healthcare lack the same structure: salespeople wait for customer feedback, lawyers communicate with clients, and doctors interact with patients. If all employees took leave simultaneously, token usage might drop sharply because most agents still wait for human prompts rather than running continuously in the background. Future legal agents should process every new contract; sales agents could continuously scan customer records for outreach signals; life sciences agents could systematically read all testing and research material. Achieving this requires enterprises to reconnect and redesign processes around AI rather than simply inserting agents into existing sequences of manual steps. Adoption outside coding will therefore involve more change management, data cleanup, and workflow redesign.
https://x.com/levie/status/2086559201053294909
Garry Tan, Y Combinator CEO
Garry Tan prefers starting with problems that have already surfaced rather than building grand abstract solutions first. The entry point might be a bug, a capability gap, a false claim, a half-finished tool, or anomalous behavior in an institution. The next step is to ask what hidden mechanism makes this visible failure possible. Once the mechanism is identified, fix the root cause rather than just treating symptoms. After one fix, look for the next anomaly and repeat the cycle over the long term.
https://x.com/garrytan/status/2086615082163941460
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari used his experience with Fable to illustrate how excessive reliance on feature flags makes product defaults hard to understand. Fable put so many features behind environment-variable-controlled flags that he had to explicitly write in Claude.md: "Defaults matter. Don't be ambiguous." This points to a specific risk in agent programming: to preserve every possibility, models may keep adding switches instead of making clear product decisions. He is also looking for genuinely effective AI collaboration experiences for multiple people. Common interfaces still primarily support one human working with one agent, while collaboration among multiple humans and multiple agents has yet to develop a mature pattern. His open question is whether this stagnation reflects a lack of product imagination or continuing limits in model capabilities.
https://x.com/nikunj/status/2086492103945900437
https://x.com/nikunj/status/2086438339419496449
Aditya Agarwal, SPC General Partner
Aditya Agarwal used Wittgenstein's two views of language as an analogy for shifts in AI research. In 1921, Wittgenstein believed language necessarily had a deep logical structure. Approximately 30 years later, he argued instead that we should stop searching for hidden structure and observe how language is actually used. Similarly, AI research in the 1960s believed intelligence rested on an explicitly expressible symbolic structure. Approximately 60 years later, the leading breakthroughs came from scaling neural nets, shifting emphasis from prescribing intelligence's structure in advance to letting models learn from usage and data.
https://x.com/adityaag/status/2086592574534602781
Official Blog
Claude Code now supports artifacts
Claude Code can now capture work in progress as an artifact, turning session content into a live, shareable visual web page. Formats include PR walkthroughs, system explanations, filterable dashboards, incident investigation pages, and automatically updated release checklists. Artifacts use the full context of the current session, including the codebase, connectors, and conversation, so they can combine failing tests, relevant functions, error spikes from monitoring tools, and root-cause analysis on one page. Users need not connect additional data sources or deploy infrastructure; they simply ask Claude Code to create a page. When an artifact changes, an already open page refreshes in place, and each publication creates a new version at the same URL. The system also provides version history, restoration, and a gallery for managing existing artifacts. During debugging, the page can accumulate timelines, suspect commits, and error-rate charts as the investigation progresses, giving engineers and stakeholders the same context. Each artifact is initially visible only to its author; sharing allows access only to authenticated organization members, and public publication is unavailable. Administrators can manage access and auditing through organization-level switches, role scopes, retention policies, and a compliance API. The feature is currently in beta for Claude Team and Enterprise organizations, with creation from the Claude Code CLI and desktop app and viewing in the browser.
https://claude.com/blog/artifacts-in-claude-code
Podcasts
Unsupervised Learning — Ep 92: xAI Co-Founder Unpacks the Future of Model Development
Key takeaway: The next stage for agents is not simply better coding. It is extending successful methods from verifiable tasks into scientific discovery, personal AI, and real-world feedback, while enabling more people to own the value and control AI brings.
Igor Babushkin is now founding River AI. Previously, he worked on StarCraft and AlphaCode at DeepMind, participated in early reasoning research at OpenAI, then co-founded xAI and worked on Colossus and model development. He sees November to December last year as a major turning point in software engineering, when a new generation of coding agents became too capable for researchers and engineers to ignore. Many personal projects that had been waiting suddenly became feasible, but rapidly growing capabilities also made developers resemble the "sorcerer's apprentice": gaining power while facing the possibility of losing control of it.
Coding and mathematics remain the most natural breakthroughs for agents because results can be verified. Code can be checked through tests, while mathematical theorems and proofs can be formalized with tools such as Lean, providing clear reward signals. Scientific discovery presents a different challenge: when agents design new materials, investigate fundamental physics, or improve rocket engines, real experiments must establish whether their proposals work. The true bottleneck is therefore not just model reasoning but closing the data-feedback loop between models and the physical world.
Personal AI need not prove the Riemann hypothesis, but it must understand users' lives, work, and environments and act on their behalf. Igor believes the current pattern of "large-scale pretraining, offering an API, and charging per token" is merely a set of industry assumptions, not the only way to build and distribute AI. River AI is exploring new approaches to model building, customization, and delivery to give businesses and consumers more direct control over AI behavior and their experience. He emphasized: "Distributing the benefits of AI and control over AI more broadly is itself part of AI safety."
He sees the more urgent risk as AI amplifying inequality, with some people gaining enormously and others left behind, rather than a distant model takeover. Open models are especially valuable when approaching dangerous capability thresholds without yet being capable of real harm, because researchers worldwide can use them to test alignment, control, and cybersecurity approaches. AI safety cannot be researched only by a small group inside large labs. For builders, what matters is not just pursuing maximum capability, but designing technologies and products that distribute value, broaden participation, and preserve ordinary people's control.