AI Builders
Digest

Anthropic launched Claude Opus 5 at the same price as Opus 4.8. It is the default on Claude Max and the strongest option on Claude Pro, with a Fast mode roughly 2.5× quicker. Official posts call it the most aligned model yet—lowest rates of reckless or deceptive behavior and strongest adherence to Claude’s Constitution. It improves on cybersecurity tasks versus 4.8 but remains substantially behind Mythos 5 at developing exploits. Claude Code’s Thariq shared that the team removed ~80% of the system prompt for the new models and published lessons on writing system prompts, skills, and Claude.MDs. Boris Cherny highlighted that Opus 5 is Anthropic’s least prompt-injectable model; combined with alignment, probes, and Claude Code Auto Mode, successful attacks drop to near zero. Alex Albert noted major gains in token efficiency while raising the intelligence bar, making it preferable for many coding tasks. Box CEO Aaron Levie reported meaningful lifts on Box’s Complex Work Eval (due diligence +17%, life sciences +30%, legal +12%). Dan Shipper’s early vibe check was more cautious: the model argues with instructions and stops early unless workflows are rebuilt from scratch, positioning it as a “poor man’s Fable.”

Open-weights models received unusually broad support. Sam Altman stated he wants the US to win in both open-source and proprietary models. Box CEO Aaron Levie detailed why open weights matter: they enable post-training for specific verticals, variance in safety and cyber approaches, more efficient training methods under compute constraints, and different cost structures for different workloads. He stressed that open versus closed is not zero-sum. Replit CEO Amjad Masad publicly asked whether Anthropic would sign the related letter and urged employees to clarify leadership’s position. Madhu Guru argued that the biggest near-term opportunity lies in adapting messy real-world workflows to foundation models—understanding how work gets done, designing evals, post-training, and building continuous feedback loops—a skillset still concentrated in a handful of labs.

The Takeaway: Scaling autonomous delivery is less about whether autonomy is possible and more about starting from the real customer use case and leveraging proprietary delivery data to solve operations, hardware, and multimodal fleet problems together.

DoorDash co-founders Andy Fang and Stanley Tang described an eight-year robotics journey on No Priors. They began exploring autonomy in 2018 as a skunkworks project (Stanley plus half an engineer’s time). Early partnerships with sidewalk robots and robotaxis confirmed the technology was a “when, not if.” The decisive lesson: most robotics startups build technology first and retrofit a use case; DoorDash starts from the customer problem and works backward. Sidewalk robots are too slow for the typical 3–5 mile delivery; robotaxis are over-engineered for a few burritos and cannot solve the last 100 feet. So they built DOT in-house—a 300-pound vehicle that reaches 20 mph and can use bike lanes and roads. It has been running fully autonomous Level 4 deliveries in Phoenix for nearly two years.

Andy shared early agentic-commerce results: 50% of restaurant trajectories via Ask DoorDash are from places the user has never ordered before, and grocery basket sizes are ~40% larger. They also launched a DoorDash CLI that lets agents restock shelves based on camera feeds. Stanley’s long-term prediction: in ten years there will be more Dashers, not fewer, because growth is so fast that a multimodal fleet (robots, drones, humans) will be required. The data moat is decisive: “We have 10 billion deliveries of data that exists nowhere else.”

Anthropic Engineering published “How we contain Claude across products,” detailing the isolation architectures for claude.ai, Claude Code, and Claude Cowork. The core principle is to enforce hard boundaries at the environment layer (sandboxes, VMs, egress controls) first, then steer behavior at the model layer. Risks fall into user misuse, model misbehavior, and external attackers. claude.ai runs code in ephemeral gVisor containers with minimal blast radius; Claude Code uses a human-in-the-loop OS sandbox that cut permission prompts by 84%; Claude Cowork runs inside a full local VM so credentials never enter the guest. The post recounts real incidents: project config executing before the trust prompt, a phishing prompt that exfiltrated credentials 24/25 times, and data exfiltration via an allow-listed domain (api.anthropic.com Files API). The recurring lesson: custom proxies and allow-lists are usually the weakest link; battle-tested hypervisors and seccomp hold up.

Google VP Josh Woodward announced Gemini Spark is live for all US Google AI Pro subscribers: drop in a school calendar PDF and have Gemini add every “No School” day to Google Calendar. OpenAI’s Thibault Sottiaux confirmed ChatGPT Work is available globally on all paid plans across mobile, web and desktop—“puts a jetpack on your ChatGPT.” Swyx shipped customizable skins and spritesheet animations for SmolForge. Vercel CEO Guillermo Rauch tested and endorsed Figma2React while asking “Where’s the ambition!” Peter Steinberger hit a new record of 66 autoreview rounds on a difficult refactor.

Garry Tan cited research showing that the speed of technology adoption over the last 200 years accounts for at least 25% of why some nations are rich today, and warned that macro productivity gains require managers to green-light radically different staffing and workflow plans—likely a 10-year rather than 2-year process. Matt Turck noted the surge in model routing: Stripe is rumored to acquire OpenRouter for $10B, while Cursor and Runway also launched routers. Zara Zhang argued the #1 missing feature is speed—intelligence is already good enough, but 1–5 minute waits create the worst ADHD window. Peter Yang agreed pure software is increasingly hard for indie developers to monetize without an accompanying service layer. Amjad Masad expressed surprise that VCs passed on Etched early and teased that anyone who hasn’t used Replit lately is in for a big surprise.