AI Builders
Digest

The Takeaway: Current LLMs freeze their weights after training and stop truly learning; real intelligence requires continual learning from an agent's own experience in a vast world, free from human-curated or synthetic-data bottlenecks.

Reinforcement-learning pioneer and author of “The Bitter Lesson” Rich Sutton, together with former student Khurram Javed, founded Oak Lab to pursue continual deep learning and abstraction formation as outlined in the Alberta Plan. Sutton insists: “I’m not weird. The field is weird. Before the AI craziness you wouldn’t have to say continual learning because it wouldn’t make any sense to talk about learning that wasn’t continual. All learning is continual.”

The Bitter Lesson’s essence is to avoid distraction by human knowledge and focus on methods that scale with computation—search and learning. LLMs are both a positive example (massive scaling on internet data) and a negative one (the internet is finite; the world is vastly larger, so reliance on human knowledge eventually holds progress back). Synthetic data remains bottlenecked by humans who must decide what constitutes good data. The real path is agents that learn models from their own experience, form the right abstractions for planning, and update weights continuously—using techniques such as per-weight step-size optimization and continual backprop to avoid catastrophic forgetting.

Oak Lab’s ambition is a system that spans low-level sensorimotor knowledge to high-level abstractions while remaining self-consistent. With algorithmic and hardware progress they expect this to become practical within 5–10 years. LLMs represent a major breakthrough in fluid language use, yet that is only about a quarter of intelligence; the rest remains unfinished.

OpenAI CEO Sam Altman announced that the company has paused some frontier reinforcement-learning training to ensure alignment, security and monitoring standards keep pace with the new level of capabilities. Model progress is now extremely rapid, and OpenAI has always said it would act if capabilities outstripped safety and alignment work. Altman stressed that the company cares deeply about AI safety, believes the field will eventually need shared safety standards, yet will act unilaterally in the meantime. Confidence in safety will increasingly set the pace of AI progress. The company remains optimistic about its alignment work and committed to making frontier capabilities widely available. Further-out releases are affected, but great new models are still expected soon.

Anthropic Engineering details how the company contains Claude’s blast radius across claude.ai, Claude Code and Claude Cowork through environment-layer isolation. Risks fall into user misuse, model misbehavior and external attacks. Defenses cover the environment (sandboxes, VMs, egress controls), the model layer (prompts, classifiers, training) and external content.

claude.ai runs code in ephemeral gVisor containers with minimal blast radius; Claude Code uses a human-in-the-loop OS sandbox (Seatbelt / bubblewrap) that cut permission prompts by 84%; Claude Cowork runs inside a full local VM so credentials never enter the guest and only the mounted workspace is reachable. The post recounts real incidents—project configs executing before the trust prompt, phishing that caused credential exfiltration, and data leaving via an allow-listed Anthropic Files API—and stresses that custom components are often the weakest link while mature hypervisors and seccomp have held up. The guiding principle is to design hard environmental bounds first, then match isolation strength to the user’s capacity for oversight.

OpenAI Codex lead Thibault Sottiaux recapped recent changes that further reduce the risk of destructive actions by GPT-5.6 inside Codex. Earlier reports showed rare cases where cleanup of temporary folders could delete user files—often by reusing system variables such as $HOME or failing to check the target path. Protections now include explicit instructions to verify deletion targets, create fresh temporary directories, avoid repurposing system variables and prefer recoverable actions; stronger execution checks that escalate high-risk commands; harder-to-enable Full access with clearer warnings; improved Auto-review; and new replay evaluations plus RL tasks focused on these failure modes. The changes substantially cut the risky behavior while preserving normal coding performance. Users should keep the app updated and prefer “Ask for approval” or “Approve for me” sandbox modes.

Anthropic Claude Code team member Thariq observed that there is an obvious “make a lot of money” button few people are pressing: take an existing SaaS product, make it headless so agents can use it directly, and charge per interaction—especially attractive for enterprise customers. As agent usage grows, this usage-based model can meaningfully expand revenue.

Box CEO Aaron Levie noted that case study after case study shows the value that can be created between the AI model and the ultimate end-user workflow is far larger than many assumed. Model capability does heavy lifting, yet diffusing AI into the enterprise still requires substantial work: domain-tuned harnesses and product experiences, connections to heterogeneous enterprise systems and data, domain-specific change management, multi-model support for cost/performance trade-offs, domain-specific evals, and pricing abstractions that match industry consumption models. Together these form a huge surface area for sustainable innovation and differentiation in the applied AI layer.

Claude Blog announced that Claude Code now supports Artifacts. Work from a session can be captured as live, shareable visual pages—PR walkthroughs, system explainers, dashboards, release checklists and more—that update themselves as the session progresses. Pages are built from the full session context (codebase, connectors, conversation) with no extra data plumbing required. Every publish creates a new version at the same link, with restore history. Artifacts are private to the author by default and shareable only within the authenticated organization; admins control access and retention. Available in beta for Claude Team and Enterprise orgs.

The official Claude account announced that Claude can now send emails in Gmail and manage files in Google Drive. Users can ask Claude to reply to a thread; it drafts and sends the response, with user control over when approval is required. Connect Gmail or Google Drive from the connectors menu. Available on all paid plans. Claude Cowork is also now available on mobile and web for all paid plans.

Vercel CEO Guillermo Rauch described the experimental open-source, model-agnostic coding CLI he now uses as his daily driver: 10–20× smaller than major coding CLIs, instantaneous startup, more like zsh than an IDE in the terminal, and embeddable even in the browser via WebAssembly. He also argued that a company’s software factory should be a monorepo containing all context—design, marketing, sales, engineering, support—so agents can build on it. Separately, Vercel is putting $1 M toward publicly verifying the security of Vercel Sandbox, inviting any model to try to escape.

Meta AI Senior Director Madhu Guru advised treating evals the way one treats frontier models: establish the quality frontier first, then work down the cost curve. Write the rubric so “good” is clear, then measure with the highest-quality method available (humans, expensive LLM judges, or automated verification) and spend the money needed for a trustworthy signal. Only after the eval reliably distinguishes good from bad and reflects product priorities should teams focus on cost—more automation, smaller judges, sampling and deterministic checks. Quality first, cost next.

Swyx (affiliated with smol.ai and others) shared that the AI Engineer team has been running extensive A/B tests on YouTube thumbnails and is open-sourcing the learnings. The goal is to help good educational content rise above the noise and to invite others to share their own experiences.

Google Labs announced that its experimental AI productivity agent CC has opened a waitlist in Australia and New Zealand and is expanding invitations in the US and Canada. CC has also been upgraded to manage calendars: once connected to Gmail it automatically creates and keeps events up to date in a dedicated Google Calendar.

Claude Code team member Boris Cherny noted that slow startup makes the Desktop app feel sluggish for daily users, and the team continues to ship small quality-of-life improvements, including further startup-speed work.

Peter Yang observed that AI has largely landed on top of existing work rather than replacing it: teams spend more time chatting with AI and delegating to agents yet have not reduced time on prior work, because expectations for what each function can accomplish have risen. At the same time non-engineers are shipping more code—PMs attaching PRs rose from 3% to 10%, designers from 1% to 8%, and founders sit second only to engineers at 23%.

Peter Steinberger celebrated receiving Apple 512 GB RAM Studios and noted the irony of having once been among those who listed reasons why certain CLI approaches could not work—until he saw the results and changed his mind.