AI Engineer organizer Swyx
Swyx recommends an AIE NYC talk in which Addy Osmani presents Anthropic’s approach to roughly 30k code changes a day. SonarSource’s Kapmani contributes research on why each loop needs independent verification. Swyx discloses that Sonar also sponsors the event. A separate announcement introduces Raising your Ambition, the final theme in AI Engineer’s year-end lineup, with Jason Liu, Thariq, Jarred Sumner, Sourcegraph’s sqs, Armin Ronacher, Dex Horthy and StrongDM Dark Factory’s Justin among the speakers. A keynote by echelon will cover cloning the Adobe suite and what comes next. The event plans to gather 1000 agentic engineers in SF and stream on YouTube to an online audience Swyx describes as 2M MAU, with registration linked in the post.
Sam Altman
Sam Altman says people should always control the future. OpenAI rejects relegating humans to doing only things that AI systems consider worth rewarding them for. He argues that people should not cede the future to superintelligence, and that this requirement should ultimately constrain AI progress.
Vercel CEO Guillermo Rauch
Guillermo Rauch recommends doctor subcommands in CLIs, using vercel doctor as an example. He says Vercel’s internal security agent automatically audits every line of code and that users can now enable Vercel Security Review on their own repositories, whether or not they host on Vercel. He argues that installing a skill, asking a model to “please make secure,” or hoping its initial output is safe is insufficient. The product was developed entirely by Vercel’s security team, and he claims it catches issues other agents miss. A separate post pairs Vercel CI’s repo acceleration with Security Review’s protection; these efficacy claims are his product description.
FPV Ventures partner Nikunj Kothari
FPV Ventures partner Nikunj Kothari recommends two relatively underused default Claude Code skills, /insights and /playground. He schedules /insights every two weeks to suggest setup upgrades, combining its findings with X bookmarks about software factories to improve how he works. He uses /playground to visualize unfamiliar GitHub architectures or simulate a complete learning loop for a side project. His latest approach pairs Codex and Claude Code with the skill to find dead code, hot paths and latency improvements. These are descriptions of his workflow, without a measured improvement figure.
AI tutorial author Peter Yang
Peter Yang observes that talented people increasingly seem to leave AI-native companies after 1–2 years or less, whereas leaving before 2–3 years once risked a job-hopper label. He acknowledges many reasons for departures while questioning how sustainable these jobs are, particularly for parents. Separately, he presents five uses to try with Grok Bot’s calling feature. For bill negotiation, supply a current $84 monthly bill and a competitor’s $49 offer; for gym cancellation, instruct it to reject retention offers and obtain a confirmation number. For a canceled flight, have it hold until reaching a person and request a refund, while a restaurant booking can specify four people at 7pm Friday with 6–8pm alternatives. It can also call three contractors for water-heater replacement prices and earliest start dates, then organize the answers in a table. Yang notes that the feature only supported US numbers at the time and links a demonstration with a real call.
Box CEO Aaron Levie
Box CEO Aaron Levie argues that AI output differences may be at their widest even as adoption is at its broadest. A few years ago, he says, skilled users in well-suited fields such as coding might have achieved 30–50% productivity gains while spending considerable time correcting errors. Limited context windows constrained long tasks, and interfaces required synchronous interaction. He argues that models now tackle longer tasks with less human involvement, supported by stronger harnesses, better domain training data and systems for managing multiple agents. He estimates that the upper bound in some fields has moved from 30–50% to 5–10X, potentially creating an order-of-magnitude output gap between frontier users and people paying no attention. He predicts RSI will widen that gap further; the figures and forecast are his assessment.
Anthropic’s Boris Cherny
Boris Cherny says Claude can automatically connect to company plugins, reducing the hassle of signing in to each separately. Team and Enterprise administrators configure them first, after which teammates connect automatically on first use.
FirstMark investor Matt Turck
FirstMark investor Matt Turck describes SurrealDB 4 as a milestone in the database’s maturation. Even after an initial open-source surge, he argues, building a strong database takes years of unglamorous engineering. He describes version 4 as combining vector, graph and document models, ACID compliance and a rebuilt engine. He also cites native Postgres/MySQL support and large-scale expansion through Cloudflare. The claimed peak is 10B+ rows/sec, positioning it as a database for the agentic era; his post does not provide full benchmark conditions.
OpenAI’s Thibault Sottiaux
OpenAI’s Thibault Sottiaux says agents remain at an early stage: impressive, yet still deeply frustrating. He wants to resolve those problems and regain a flow state, while recognizing a long road ahead.
Former Claude Code contributor Thariq
Thariq, formerly of Claude Code, frames a prediction as the bitter lesson at work: smarter models will need fewer tokens. He predicts they will no longer need to verify their work or check for bugs because they will simply be right; this is a forecast, not demonstrated present-day reliability.
Replit CEO Amjad Masad
Replit CEO Amjad Masad says enterprise adoption of agents and vibe coding is deep in a chaotic phase, without detailing organizational cases or implementation methods in the post.
Google Labs
Google Labs says Dreambeans is reaching new users in India and invites their feedback. It asks which region should come next, without specifying a further rollout list or date.
Cat Wu, departing the Claude Code and Cowork teams
Cat Wu announces that the previous Friday was her last day at Anthropic. Two years earlier, when Claude Code was a small internal tool, she asked it to build an RL environment, steered it for tens of hours without manually writing code, and joined the team after it worked. At launch, most users approved each permission prompt and mainly used Claude for answers and tests. She says users now more often trust Claude’s judgment with automatic approval, while it tackles large migrations and new products. She credits thousands of team ideas behind Claude Code, Cowork and Tag and thanks users whose feedback shaped the products. Her account describes changing usage habits rather than establishing that every task should receive automatic authorization.
Unsupervised Learning — Ep 96: Nebius CTO on Racing the Hyperscalers Before AWS Wakes Up
Key takeaway: An AI cloud’s differentiation rests on software, execution and reliable agent outcomes as well as the number of GPUs.
Nebius CTO Danila Shtan starts with the infrastructure experience of the team that came from Yandex. After the 2024 separation, the software platform was rewritten from scratch. He recalls roughly 40–50 people handling GPU procurement, servers, firmware and networking while the software organization already numbered about 200. The goal was a single platform serving both clusters with tens of thousands of GPUs and a customer renting one GPU for half an hour.
Deploying new hardware requires repeated joint debugging. NVIDIA’s prototypes, designs and samples keep changing, while the cloud provider must integrate virtualization isolation, storage and networking and report problems back. Moving from individual servers to rack-scale architectures such as GB300 also changes maintenance and workload migration. Shtan sees execution across simultaneous constraints as decisive: equipment, power, components and construction can all be scarce.
Customer differences often arise from maturity and compliance obligations. He says new labs tend to prioritize performance, capacity and availability, while established enterprises can demand certifications, security checks, regional compliance and code review even on smaller contracts. Storage must be judged beyond capacity, too. He cites customers seeking 500–700 GB per second from one shared volume as an example of scale his team initially underestimated.
Inference optimization starts with actual traffic and tasks. Shtan describes work on request routing, inference-server configuration, kernels and model compression. Prefill/decode disaggregation is not the default for every customer; some demanding workloads use manually tuned deployments. For bounded-domain agents, he emphasizes predictable outcomes and elapsed time. He says, “Whatever you’ve done with the agent, it’s still your result. It’s your responsibility. It’s your pull request.” The team does not impose individual usage budgets but monitors ineffective use and keeps fully autonomous agents away from core services for now.
He is also interested in optimization outside the model: tools, harnesses, quality gates and CPU-side code can advance work for minutes without another LLM request. Demand growth, hyperscaler competition and the survival of smaller providers remain his forecasts. His current research interest is how workflows reliably reach their intended outcomes.