Garry Tan is reworking Geoffrey Moore's Crossing the Chasm for a new kind of market. His argument: Moore's framework assumes there's an existing solution the buyer is comparing you to. But when the alternative is literally nothing—that is, the bar is zero—the chasm doesn't exist.

Buyers start acting like visionaries because they have to buy. They'll tolerate a 60% solution because 60% of something beats 100% of nothing.

The practical takeaway for founders: if you're in a bar-is-zero market, stop worrying about whole product, references, and pragmatist buyers. Ship the 60% solution and iterate from there.

Aaron Levie is calling a Jevons paradox for security engineering. AI makes it far easier to find security issues, which means the new bottleneck is reviewing, triaging, and fixing them.

Far from AI eliminating security engineers, the flood of newly discoverable vulnerabilities creates more demand for human judgment and follow-through.

His one-liner captures it: 'We're about to enter a security engineer boom.'

Swyx is co-signing a framework for what kinds of learning transformers do well today and where they hit limitations.

His key point: throwing more params and more power at a demonstrably inefficient paradigm will eventually be outclassed by a simpler solution that can hypothesize and seek truth rather than backfit a house of cards.

But he also acknowledges the bitter lesson—it may be simpler to just scale, and we may hit AGI anyway because human intelligence just isn't that smart nor plentiful.

Oriol Vinyals, Gemini co-lead, gives the most grounded view of where model research actually stands. His key insight: the 'GPT moment' for video and images hasn't happened yet. Current multimodal recipes mix everything together and get good results, but the pure transfer from visual data to conceptual understanding—without language as a crutch—is still unsolved.

On continual learning, he's betting on a file-system-style nonparametric approach: models write their thoughts to files, structure them in directories, and read them back later. This is more practical than integrating memories into weights, which creates a nightmare for serving one model at scale to different users.

On innovation itself, he's honest: he hasn't seen truly outstanding ideas generated by models yet. The models understand mechanistic details of training superhumanly, but the leap to genuinely novel ideas—especially in machine learning itself—hasn't happened.

Swyx is prototyping a new pattern for AI engineering: Kakuna, a system of skills with checklists that harden your codebase. The idea is you /plan with it, then let it /goal for a day, and it comes back with the same functionality but all the boring hardening work done—plus an audit of its own work.

His framing is memorable: instead of a 'dark factory,' go 'mullet factory'—party in front (ship unique lovable features), dark in the back (timeless production principles). It's the anti-entropy, anti-slop part of the development cycle, broken out as its own thing.

The deeper point is about subagent parallelism and encoding strong opinions on how AI engineers should design apps for both human and agent access, DevOps, and product management.

Peter Yang's list is notable because it's not about resisting AI—it's about building the kind of leverage and optionality that make you resilient regardless of what happens.

His most concrete advice: learn Codex or Claude Code as training grounds for working with AI agents. Build side projects to revive atrophied builder skills. Develop a GitHub history so your work is visible. Become top 10% at one craft, because AI gets people to average fast, which means customers pay more for genuine human craft and taste.

His closing line is provocative: 'Entrepreneurship is the safest job in the AI era anyway.'