What the people actually building AI said today. One page — a 4-min read.
Price was the headline: OpenAI cut GPT-5.6 rates by up to 80% and framed it as owning the price/intelligence tradeoff at every tier. Underneath that, a security incident had half the timeline arguing about what agents do when they hit a misconfigured system, with the people who actually run sandboxes for a living pointing out that the mistakes are mundane, not sci-fi. Plus a good reminder that the largest AI deployment in the physical world belongs to a company that sells dash cams.
X
Sam Altman
OpenAI cuts GPT-5.6 prices up to 80%, adds Fast mode to the API
GPT-5.6 Luna drops 80% to $0.20 per million input tokens and $1.20 output. Terra drops 20% to $2/$12. Sol gets Fast mode in the API: up to 2.5x the speed for 2x the price with the same intelligence. Altman's framing for all of it is that OpenAI wants the best price/intelligence tradeoff at every level, not just at the frontier.
The agent security lesson is about hardening enterprises, not fearing AI
Given the right tools and a task, agents will do whatever it takes to finish the job if they have enough compute, so anything misconfigured or only assumed to be locked down becomes a risk vector. Levie's read is that the real implication is not AI trust and safety but the volume of work enterprises now owe their own environments. Separately he laid out the cost cycle he expects to repeat indefinitely: frontier models arrive capable and expensive, then efficiency gains or competitive pricing drop the cost per task for everything below the peak, and cheaper tasks pull in wider adoption.
Replit on sandboxes: assume zero-days exist, because they do
The wave of AI escaping sandbox stories reads as scary AI, but Masad's diagnosis is that most AI companies and newer sandbox providers are making very basic mistakes. Replit has run sandboxes since 2016 and been targeted by every hacker and state actor around, and the advice that came out of it is layered protection inside a zero-trust framework rather than any single boundary. Assume the zero-days are already there.
Grok Build apps run on Vercel, and CLI-to-live-URL got ~7s faster
Apps built by prompting Grok and published to *.grok.me are backed by Vercel hosting and CDN. Vercel also shaved up to roughly 7 seconds off the end-to-end CLI to live URL deploy path for many apps, which matters most when the thing driving the CLI is an agent rather than a person. Rauch is explicitly courting anyone building a platform that writes and ships software autonomously on top of Vercel's CLI, MCP, and API surface.
Samsara: 99% of US roads daily, 25 trillion data points, $2B ARR
The moat in physical AI is data nobody can scrape. "These are not the tokens you're gonna find online. Like, you can't crawl Reddit and find out about what happened on a construction site." Samsara runs GPS trackers, asset tags, and dash cams across millions of vehicles, credits the system with preventing about 380,000 road accidents last year, and just shipped Agent Studio, where a warranty agent reads a fault code, cross-references the OEM warranty and service manual, and opens the work order in under a minute instead of an hour or two of human labor. Biswas also pushes back on the surveillance framing with a number: dash cams are used mostly for exoneration, and Home Depot saw a 65% reduction in auto claims. On autonomy he splits the forecast, robotaxis soon because regional operations look alike, commercial trucking on a ten to twenty year diffusion curve because cement mixers and garbage trucks are custom built and the long tail is messy.
Chasing pretrain data quality forces you to build a private Google
Once you decide CommonCrawl is not good enough, you have to build a whole web scraper, and keeping it current means indexing, and at that point you have accidentally built a private low-frequency clone of Google as a side project of pretraining. The payoff is that the same infrastructure gets reused for agent inference. Labs do use third-party search providers, but swyx reads the drift toward first-party equivalents as both a competitive advantage and a target for AEO mimicry, which is exactly why none of them will talk about it. He also noted that if you can distill models, you can distill agent harnesses.
Want to train a nontechnical team on AI? Run an install party
Setup is 80% of the barrier, so the fix is to get everyone in a room with their laptops, install the agents on their machines right there, and have them complete one meaningful task on the spot. Skip the abstract talk entirely. Once it is installed, people start talking to the agent and learning from each other without further instruction.
How you'll know the models got really good: reliability rising under load
The tell will not be a benchmark. It will be reliability climbing while load keeps going up, sudden efficiency gains, things getting faster, resets. Sottiaux is also openly soliciting Codex improvements, with no ask too small.
GCC now flatly rejects LLM-generated code, with no way to detect it
The policy change bans LLM-based contributions outright. Steinberger's objection is enforcement: there is no mechanism to prove a patch came from a model, which makes the rule theater.
Preseen's risk forecasting is in private beta with quant and hedge funds
A number of the world's leading quant and hedge funds have been using it over the last few weeks specifically to avoid getting blown up. Agarwal tied the pitch to the week's news, arguing the tool would have caught the risk that surprised Leopold.
Source data comes from the open-source project follow-builders by zarazhangrui, released under the MIT license. Summaries are generated by an LLM from that project's public feeds, and the summarization prompts are adapted from it. Every item above links to its original source.
Summaries generated automatically. Read the original before relying on any claim.
