X / Twitter
Swyx, AI Practitioner
Swyx proposed that if models can be distilled, agent harnesses—the systems governing how agents run, call tools, and control workflows—should be distillable too. This extends distillation beyond model capabilities to the external execution framework. He also noted that if a lab’s standards for pretraining data quality are too high to accept Common Crawl, it must build its own system for crawling the entire internet. To keep obtaining fresh data, that system also needs indexing capabilities, eventually becoming a completely private Google clone updated at a lower frequency. It may begin as a pretraining side project, then directly serve agents’ search needs at inference time. Swyx believes that continuously developing first-party search capabilities creates a competitive advantage but also makes them a target for AEO imitation attacks, so labs will be reluctant to disclose them.
https://x.com/swyx/status/2083073422410821846
https://x.com/swyx/status/2083016652032188669
Thibault Sottiaux, OpenAI Codex and ChatGPT
Thibault Sottiaux described several outward signs that a truly excellent model might leave. Even as system load continues to grow, reliability would improve rather than deteriorate with scale. Sudden efficiency gains, faster responses, and system resets might also reflect changes in underlying capabilities. He also publicly solicited issues affecting everyday Codex use, stressing that even the smallest improvement suggestions are worth raising. Together, the two posts point to a product judgment: model progress shows up not only in benchmarks but also in reliability, efficiency, and improvements to details ordinary users notice every day.
https://x.com/thsottiaux/status/2083053369351090254
https://x.com/thsottiaux/status/2083048892405604681
Amjad Masad, Replit CEO
Amjad Masad emphasized that the real difficulty is getting sandboxes right, rather than simply attributing every escape to AI being frightening. He argued that many AI companies and recently emerged sandbox providers are still making very basic security mistakes. Replit has operated sandboxes since 2016 and has long faced targeted attacks from hackers and state-level adversaries, accumulating practical defensive experience. His first recommendation is to assume that zero-day vulnerabilities exist, because in reality they do. Security design cannot rely on a single isolation boundary; it needs multiple layers of protection within a zero-trust framework. This shifts the focus from whether models are dangerous to whether infrastructure is designed on the assumption of ongoing compromise.
https://x.com/amasad/status/2083034412598579403
Guillermo Rauch, Vercel CEO
Guillermo Rauch said Vercel had reduced end-to-end deployment time from CLI to Live URL by up to approximately 7 seconds for many applications. The optimization affects not only manual deployments but also agents that build and release software automatically. Developers can access Vercel’s infrastructure through the CLI, MCP, or API and build custom software factories on top of it. He specifically invited teams building agents or platforms that autonomously develop and deploy software to contact him. Meanwhile, `*.grok.me` applications generated by Grok Build are now served by Vercel’s hosting and CDN infrastructure. Users can generate games, websites, internal applications, or personal software through prompts, then use Publish to release the product to one user or a billion users. Together, these developments shorten the delivery path from natural-language intent to a live application.
https://x.com/rauchg/status/2082876367629381719
https://x.com/rauchg/status/2082841035093467229
Aaron Levie, Box CEO
Aaron Levie argued that the lesson from the security incidents should not be that AI is frightening, but that security configurations must be done properly in the agent era. Given suitable tools, a clear task, and enough compute, an agent will continually search for ways to achieve its goal. Misconfigured systems, or systems mistakenly believed to be locked down, can therefore become entry points for risk. The problem to solve extends beyond the model’s own trust and safety to comprehensive hardening of enterprise environments. He expects most organizations still have substantial practical work to do in this area. On economic diffusion, he argued that falling AI costs normalized by task type are key to broadening adoption. Frontier models may appear expensive because they handle more advanced tasks, but subsequent efficiency gains or stronger competition lower the price of equivalent tasks. This cycle repeats, making AI affordable for more use cases and driving its expansion throughout the economy.
https://x.com/levie/status/2082997703458570412
https://x.com/levie/status/2082911418349920617
Matt Turck, VC at FirstMark Capital
Matt Turck described Samsara as an enormous but rarely discussed physical AI deployment. Its systems cover 99% of US roads daily, giving AI access to real-world environmental data that cannot be obtained from public webpages in the software world. The business has reached $2 billion in ARR, processes 25 trillion data points, and is associated with the prevention of approximately 380,000 traffic accidents. Its technology stack begins with sensors, vehicle gateways, and cameras, extending to edge AI, cloud AI, video reasoning, and real-time driving coaching. Samsara Agent Studio also applies agents to practical workflows such as warranties, controlling risks through rules, workflows, and guardrails. The discussion also covers the boundaries of workplace AI surveillance, how cameras can protect drivers or establish that they were not at fault, and the development of autonomous trucks and mixed human-machine fleets. The full content is available on Spotify, Apple Podcasts, and YouTube.
https://x.com/mattturck/status/2082907703735730609
https://x.com/mattturck/status/2082907699646173484
Zara Zhang, Builder
Zara Zhang’s AI training advice to managers is always the same: organize an install party. Everyone brings their own computer and installs an agent on their device on the spot. Once installation is complete, each person must immediately use the agent to accomplish a meaningful task. She recommends skipping abstract introductions because explaining concepts cannot replace a first real experience. Once the tool is in their personal work environment, people naturally start interacting with the agent and learning from how colleagues use it. She estimates that 80% of the barriers to adoption come from installation and initial setup. For nontechnical teams, the priority in training is therefore to get everyone past that first hands-on hurdle, rather than teach more AI concepts.
https://x.com/zarazhangrui/status/2083084770763002350
Sam Altman, AI Observer
Sam Altman announced several model pricing adjustments. GPT-5.6 Luna’s price has fallen 80%, to $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra’s price has fallen 20%, with input and output priced at $2 and $12 per million tokens respectively. GPT-5.6 Sol has added Fast mode in the API, offering up to 2.5 times the speed at the same intelligence level for 2 times the price. This pricing structure addresses both lower costs for routine inference and scenarios where users are willing to pay a premium for response speed.
https://x.com/sama/status/2082880720989532597
Podcasts
The MAD Podcast with Matt Turck — The Biggest AI Deployment Nobody Talks About | Samsara CEO Sanjit Biswas
Key takeaway: The real barrier to entry in physical AI is not building another model, but first turning the undigitized real world into reliable, actionable data.
Samsara co-founder and CEO Sanjit Biswas previously pursued a PhD at MIT and founded Meraki with co-founder John out of the campus Wi-Fi research project RoofNet. Samsara emerged from a fresh learning process across industries. They had no prior experience with warehouses, loading docks, or construction sites, so they had to enter the field to understand physical operations rather than infer needs from books.
The first key asset is data unique to the real world. Samsara devices cover millions of vehicles that travel 99% of US roads daily, processing approximately 25 trillion data points annually. Biswas said: “These aren’t tokens you can find online. You can’t crawl Reddit to find out what’s happening on a construction site.” Only by combining GPS, cameras, weather, speed limits, and vehicle status does a physical environment suitable for AI reasoning emerge.
The second shift is from reporting to action. Over the past twenty years, IoT mainly collected data and produced tables, leaving humans to decide what to do next. AI can now combine multiple clues to generate insights, while agents can schedule work or execute parts of a process. Samsara estimates its systems helped prevent approximately 380,000 road accidents in the previous year, but environments involving human safety also demand stricter cybersecurity, rules, and guardrails.
The third challenge comes from hardware and frontline deployment. Devices must withstand harsh conditions and transmit data over unstable networks, while millions of frontline workers must incorporate new tools into their daily routines. Physical industries such as construction, transportation, energy, and manufacturing account for approximately 50% of global GDP, representing enormous potential value, but there is no low-hanging fruit comparable to connecting an existing database. Biswas therefore sees real-world hazards as opportunities, provided the system does not introduce new risks.
The fourth signal comes from the construction of AI infrastructure. One large energy company said it planned to triple, within the next five years, the power supply capacity built over the past 125 years, with 90% of the additional demand related to data centers. Expansion is constrained by shortages in skilled trades such as electricians. AI’s role is not simply to replace them, but to reduce waste from waiting for materials, planning tasks, and coordinating on site, allowing scarce skills to focus on work that genuinely requires human judgment.
Physical AI’s competitive advantage ultimately comes from the combination of sensors, data, models, hardware, and field workflows. The harder the real world is to crawl through the internet, the harder it is to replicate systems that can continuously observe it and act safely.