X / Twitter
Peter Yang, Practical AI Tutorial Author
Peter Yang pointed out that discussions of AI risk are intense, yet the vast majority of people worldwide have not even really started using AI. He cited a chart frequently used by Brex CEO Pedro, in which each square represents 3.2 million people: 84% have never used AI, and 16% use free chatbots. Those paying $20 per month for AI account for approximately 0.3%, while only around 0.04% can use agents effectively. The chart is reportedly from February 2026, and adoption may have increased since then, but paid usage and effective use of agents remain at a very early stage. He also reported that OpenAI's Record & Replay feature seemed unable to trigger correctly, illustrating that many product details still need to be resolved as agent tools move from broader awareness of the concept to stable workflows.
https://x.com/petergyang/status/2099200231820963891
https://x.com/petergyang/status/2099175305818890472
Amjad Masad, Replit CEO
Amjad Masad said the prices of AI coding during the preceding period had made it difficult for many users to keep using it, which he found painful. He then announced that users could now build software for free again. The central signal was not an ordinary price reduction, but a return to a zero-cost entry point for AI coding. For students, independent developers, and founders who have yet to validate demand, free access means they can finish a prototype before deciding whether to invest more. This also reflects a shift in competition among AI coding products: beyond simply improving model capabilities, the question is expanding to who can enable more people to consistently complete real projects at low cost.
https://x.com/amasad/status/2099197117013340450
Aaron Levie, Box CEO
Aaron Levie believes "pacing" can easily be understood as deliberately slowing AI capabilities and could even be used as a tool to regulate competitors, but that the specific improvement goals Dario proposed remain essential requirements for AI development. Aerospace, life sciences, and healthcare all require rigorous safety validation; as AI enters financial trading, medical devices, biotechnology, defense systems, and government workflows, it cannot be an exception. He emphasized that the need for system safety and alignment is real, and that the debate should focus on how to achieve them, rather than whether they are necessary. He acknowledged that achieving this without significantly slowing innovation or weakening market competition is one of the most complex problems of the 21st century. This view brings AI safety back from abstract ethical discussion to questions of engineering and institutions: critical systems need validation, oversight, and accountability mechanisms commensurate with the scope of their impact.
https://x.com/levie/status/2099167992835924301
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari cautioned job seekers that joining a startup with a higher valuation and seemingly higher total compensation can often be a drawback rather than an advantage. A financing valuation does not mean a company is safer; evaluating an opportunity requires an objective look at market size, business traction, and competitive position. If a company is valued at 100 times ARR, it still needs subsequent growth to justify that price, and candidates should independently estimate the price at which it might eventually exit. In addition, the 409A price and related tax implications directly affect the actual value of equity compensation, making them particularly important in a market where valuations have become detached from reality. He illustrated the risk with his own experience: he once joined a company that had raised $60 million from top investors, but it shut down three years later, leaving all the equity worthless. When evaluating a startup offer, therefore, funding news should not be treated as proof of safety; valuation, required growth, potential exit value, and tax costs should be calculated together in a single table.
https://x.com/nikunj/status/2099272635926671532
https://x.com/nikunj/status/2099198567923765357
Builder Zara Zhang
Zara Zhang pointed out a clear gap in Astra's execution after being corrected. She told Astra that the X it had already completed was wrong and that it should do Y instead. Astra immediately acknowledged the issue and said it should execute Y, but stopped at verbal confirmation without taking further action to make the change. She contrasted this behavior with other models, which typically carry out the new instructions directly after accepting a correction. The example exposes an important standard for the experience of using agent products: acknowledging a mistake is a conversational capability, while the ability to turn corrective feedback into follow-up action determines whether a task is actually completed.
https://x.com/zarazhangrui/status/2099348631291883945
Sam Altman, OpenAI
Sam Altman grouped the potentially severe consequences of AI progress into two categories: humanity losing control of the future, and excessive concentration of power in the hands of individuals, companies, labs, or countries. He stated clearly that AI must always serve people, so alignment and safety techniques need to stay ahead of growth in model capabilities. For frontier AI governance, he supports uniform US federal safety requirements and suggested independent audits as one possible mechanism. He believes early Responsible Scaling Policies and Preparedness Frameworks focused primarily on deployment after a model was complete, whereas safety development and evaluation now need to move earlier into the training process. OpenAI now establishes an explicit safety case before starting frontier reinforcement learning runs expected to produce significant capability gains, while continuing its pre-release safety work. He hopes the industry will develop shared standards for misalignment, monitoring, and safety, but also emphasized that pacing does not mean stopping progress. Safety cases and monitoring impose costs and slow some development, but he believes competitive pressure cannot justify allowing model capabilities to outpace alignment and monitoring.
https://x.com/sama/status/2099352016988614852
https://x.com/sama/status/2099348812305473766
Podcasts
No Priors — Redefining Chip Architecture with Arm CEO Rene Haas
Key takeaway: AI is not only increasing demand for GPUs; it is also reaffirming the CPU's value as the coordination center of computing systems, while accelerating the reorganization of chip design, verification, and supply chain capabilities.
Rene Haas, CEO of Arm and SoftBank Group International, has spent many years in the semiconductor industry and worked at Nvidia before joining Arm. He explained that Arm originally participated in the smartphone, data center, and automotive markets primarily by licensing CPU, GPU, and system IP. Customers manufactured chips themselves or sent their designs to TSMC for tape-out. As product cycles stopped slowing down and manufacturing cycles kept lengthening, Arm first expanded from individual IP components into compute subsystems, offering customers complete combinations akin to Lego blueprints, before moving further into physical CPU products.
A direct catalyst for this shift came from Meta. Meta needed a general-purpose agentic CPU, but no supplier on the market could provide one directly, so it chose to co-develop one with Arm. Before proceeding, Arm consulted its ecosystem partners and encountered less resistance than Haas expected. More proprietary or open-source software would ultimately increase the value of the entire Arm ecosystem, and Arm server-chip participants such as Nvidia, Amazon, Microsoft, and Google also supported the direction.
Physical chips also change Arm's business model. Its previous IP licensing business had no burdens from inventory, returns, or scrapping; the episode mentioned that Arm once had a gross margin of 98.5%. Entering the product business requires the company to build capabilities in supply chain operations, back-end design, layout, implementation, and lab bring-up, and to coordinate directly with TSMC, Samsung, Micron, and SK hynix on wafers, substrates, and memory configurations. Arm remains a fabless semiconductor company and has no plans to build its own fab, but it now has to take on more of the complexity of actual manufacturing.
One of AI's most direct benefits for the chip industry is shortening design cycles that last 24 to 36 months. Haas noted that design itself is not what consumes the most time; verification, validation, and debug do, and those are precisely the areas where AI excels. He used a pointed analogy to sum up the absurdity of blocking AI tools: "It is like going back to the 1990s, when the internet already existed, but saying you could only use it from two to four each day." His view is that AI can no longer be put back in the bottle, and companies must learn how to make it a broadly used productivity tool.
Haas also rejected the idea that the AI era needs only accelerators. Training and inference produce large numbers of tokens, but systems still need to decide where those tokens go, how to schedule resources, and how to connect memory, GPUs, and other accelerators. He compared CPUs to trucks delivering tokens to users, emphasizing that CPUs, accelerators, and memory still form a complete system. This logic applies not only to data centers but also to cars, robots, phones, and wearables. Smaller AI devices that place greater emphasis on energy efficiency are precisely where Arm has an advantage.
On industrial policy, Haas supports adding more domestic fabs in the United States to protect national security and diversify supply chain risk. He also sees technological competition as an "infinite game" with no clear endpoint, in which simply restricting chip exports may not produce an ultimate winner. Nor are data centers merely enormous warehouses with only two cars in the parking lot: energy, liquid cooling, and infrastructure jobs will still develop around them. For AI builders, what deserves attention is not whether GPUs or CPUs replace one another, but how token production, system scheduling, memory, energy efficiency, and supply chains scale together as a whole.