TL;DR
Madhu Guru said Google’s AI comeback was a systems game, while Alex Albert credited Claude with helping Firefox crush 15 months of bug fixes. Peter Steinberger called GPT-5.5 refactors automatic, and Ryo Lu pushed idea-to-merge into one flow.
They say Gemini’s rise wasn’t magic — it came from building the model playbook, the customer feedback loop, and the enterprise business until all the pieces clicked with Gemini 3. That’s a pretty blunt postmortem on how Google went from behind OpenAI and Anthropic to back at the frontier. Also: they’re leaving Google, and the toddler shoutout is doing some heavy lifting.
With Claude Mythos Preview, the Firefox team fixed more security bugs in April than in the previous 15 months combined. It’s a strong signal that AI isn’t just writing code — it’s already speeding up real security work in production.
He argues that whenever AI makes a task easy, everyone else gets that same advantage fast — so the real edge moves elsewhere. For software, that means more emphasis on sales, marketing, and customer success; for finance, it’s client engagement. The core question: if everyone can do what you do, what actually makes you stand out?
He says `/goal + GPT-5.5` is now good enough to plan huge refactors with end-to-end tests and just execute. That’s the real shift: AI isn’t just helping write code, it’s starting to handle the messy orchestration work that usually slows senior engineers down.
He argues weekly 1:1s are often less about support and more about control — a way for managers to keep people in line instead of letting them do great work. The bigger point: if your job feels overly coddled, he thinks you should find a team that actually pushes you and shows you what high performance looks like.
He says the AI platform war is here, and the real battle is shifting from model quality to who owns the developer stack. In a quick dispatch from Code with Claude, he and Kieran Klaassen dig into the xAI compute deal, managed agents, and Anthropic turning its API into full cloud infrastructure for builders.
He says frontier AI gets safer through preparedness, defenses, and security work, not just bigger models. In a long convo with OpenAI board member Zico Kolter, he digs into jailbreaks, agent attack surfaces, and why the real risk is how systems are deployed, not just how they’re trained.
He dropped GStack v1.28, adding browser downloads, headed mode, anti-bot handling on headless Linux, and an llms.txt so agents can use the right skills with less guesswork. He also pushed GBrain thin-client mode, so secondary agents can lean on it over MCP instead of spinning up their own server.
Zico Kolter says AI safety won’t come from bigger models alone
The Takeaway: Frontier AI gets safer only when teams build safety into the system, not by hoping scale fixes it.
Zico Kolter, OpenAI board member and chair of its Safety and Security Committee, comes at AI safety from both sides: he runs Carnegie Mellon’s machine learning department and has spent years in AI security research. His core view is blunt: “You can’t just sort of trust models to get safer by getting bigger.” Capability scales fast; safety does not.
That’s why he pushes layered defenses instead of magical thinking. OpenAI’s preparedness framework, and similar systems at Anthropic and Google, set thresholds for risky capabilities like bio, cyber, and self-improvement. But Kolter says that’s only one slice of the problem. The real safety stack also has to cover model behavior, user misuse, and the broader ecosystem as AI becomes embedded everywhere.
He’s especially skeptical of the lazy “wait for the next model” mindset. That works for math or coding. It doesn’t work for robustness. As he puts it, “to make models more robust, to make them broadly safer, you need to be explicit in training them for safety.” In other words: safety is engineered, not inherited.
Kolter also rejects the cartoonish “doomer vs. accelerationist” framing. He sees most serious researchers in the same camp: excited about the upside, unwilling to ignore the downside. The useful question isn’t whether AI is good or bad. It’s whether the industry is building the governance, monitoring, and release discipline to keep up with what these systems can now do.
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.