OpenAI CEO Sam Altman used a long-form interview on David Senra’s Founders podcast to lay out a set of ideas that stretch well beyond product updates. He said one of the clearest signs of how early AI still is can be found in his own habits: despite helping build tools such as Codex, he still works much like he did 20 years ago, copying and pasting across messaging apps, triaging email the old way, and keeping to-do lists in familiar formats. For Altman, that is evidence that social and behavioral inertia will slow AI adoption even if the technology keeps moving fast. He also revisited the bets that shaped OpenAI, from pursuing AGI in 2015 to backing large language models after that, both of which he said were mocked at the time. Altman argued that research follows power laws just as venture investing does, with one exceptional idea outweighing everything else. On strategy, he said OpenAI should be thought of primarily as a platform company built around a unified interface and an API, and described why the company cut Sora and the Atlas browser in order to redirect compute and talent. He called compute expansion one of the most expensive infrastructure efforts in human history, and identified loss of control and concentration of power as the two AI risks he worries about most. The interview also touched on OpenAI’s four-and-a-half-year stretch without a product and on a personal habit Altman briefly kept after his first child was born: writing weekly letters that forced more honest self-reflection.

A Twitch streamer from Connecticut has filed a class action in the U.S. District Court for the Northern District of California, accusing Twitch and parent company Amazon of using creators’ livestream content to train generative AI models without authorization or payment. The complaint says the companies copied millions of videos to build commercial AI products and argues that the practice breached an implied contract with creators and violated California unfair competition law. The dispute centers on a new opt-out setting announced by Twitch on the 12th of this month. Under the policy, channels are included by default unless users manually turn the setting off. Twitch’s chief product officer, Mike Minton, drew backlash after saying on the company’s Patch Notes livestream that “if it was opt in, nobody would do it.” According to the source material, Amazon may use livestreams, VODs, clips, chat messages, images, and text for current and future models that generate or synthesize text, audio, images, or video. The complaint also argues that a channel-based opt-out system cannot secure consent from every participant and alleges Twitch had already begun collecting content without permission in 2024.

Nvidia said its general-purpose coding agent AVO achieved a perfect score on ARC-AGI-3, finishing all 183 levels across 25 game environments in 6,624 steps with an RHAE of 100.00. The result did not come from changing the base model. Instead, Nvidia wrapped Claude Opus 5 with a system layer that adds persistent memory and a supervisor, lifting performance from 30.16% for the standalone model to a full score in the benchmark setup described in the report. According to the source material, ARC-AGI-3 places agents in unfamiliar games without giving them explicit rules or goals. Some levels allow movement and rotate the entire scene when a blue-black block is touched, while others allow only clicking to cycle cell colors into a target pattern. Nvidia also noted that AVO worked entirely in text mode, receiving each frame as an exact 64×64 text grid rather than images or image tokens. The same architecture was originally built for GPU kernel optimization. A paper uploaded to arXiv on March 25, 2026, described AVO as an agentic mutation operator for autonomous evolutionary search. In tests on Nvidia’s B200, the system ran autonomously for seven days, explored more than 500 optimization directions, and produced 40 valid kernel versions. The report said its multi-head attention kernel was up to 3.5% faster than cuDNN and up to 10.5% faster than FlashAttention-4, with separate GQA results showing gains of 7.0% over cuDNN and 9.3% over FlashAttention-4 after about 30 minutes of autonomous work.

OpenAI CEO Sam Altman pushed back on the idea that artificial intelligence will remake business overnight, saying the technology may be ready before companies, customers, and institutions change their behavior. In a conversation with Founders podcast host David Senra, Altman said one of the biggest factors slowing disruption is the economy’s "huge inertia" — people keep buying from the same companies, using the same tools, and operating in familiar ways even when better technology is already available. Altman made the comments while discussing a bold prediction from Shopify CEO Toby Lütke, whom he described as one of the most forward-leaning CEOs he knows. According to Altman, Lütke believes 2026 will be the year every business gets reshuffled and that someone will build an AI-native version of Shopify. Altman, however, said he has become more cautious on timing. He recalled that when GPT-4 launched in 2023, he expected the software industry to be disrupted much faster than it has been. To illustrate his point, Altman pointed to an earlier consumer shift: even after Netflix introduced DVD-by-mail, people still walked into Blockbuster stores. For him, that was a reminder that changing habits can be harder than inventing the technology itself. He also outlined OpenAI’s product direction as a single interface connected to a personal or company AGI, paired with an API for developers.

OpenAI and Anthropic both saw fresh model clues emerge at nearly the same time, while OpenAI CEO Sam Altman used a long-form podcast appearance to challenge a strain of AI industry messaging built around catastrophic risk. According to the source article, an OpenAI employee’s public GitHub pull request briefly included the name “gpt-nathree,” adding to earlier sightings of “gpt-mewfour” in public testing tied to Codex. The article frames both names as possible checkpoints for a next-generation OpenAI agent model, with some observers speculating about a connection to GPT-6 Astra, though OpenAI has not disclosed core architectural details. On the Anthropic side, two names — “claude-marshmallow-eap” and “claude-melon-eap” — were reportedly exposed through a third-party developer app and a Discord community. The source says neither is believed to be at the Fable level, and discussion has centered on whether they are upgrades within the Claude 5 line rather than a brand-new flagship generation. Altman’s comments on David Senra’s “Founders” podcast added a second layer to the day’s developments. He said he had been wrong about how quickly GPT-4 would disrupt software markets, argued that social inertia has slowed AI adoption, and criticized what he described as a “benevolent dictator” style of AI marketing that pairs warnings about mass job loss or even civilizational destruction with promises of abundance and medical breakthroughs. He also revisited OpenAI’s early years, saying the company spent four and a half years searching in the dark before shipping a real product, and argued that in an AI-saturated future, authentic human connection will only become more valuable.

OpenAI has reduced pricing for GPT-5.6 Sol across both API usage and overage credits for the next three months, with the API-side changes already in effect. The biggest cut hit output tokens, which dropped from $30 to $20 per million tokens, while input pricing moved from $5 to $4. Subscription tiers including Pro, Plus, and Business were left unchanged, and included usage was not increased. Cache pricing and long-context pricing were also lowered. The move lands as competition around flagship AI models shifts beyond benchmark performance and into pricing. The original report framed the cut as pressure on Anthropic, which is reportedly in its IPO preparation window, while also pointing to low-cost competition from newer models such as DeepSeek V4 Flash and a model referred to as Ox Alpha. At the same time, OpenAI linked the lower prices to efficiency gains in how Sol now runs. Separately, Codex lead Tibo said Codex weekly active users have passed 20 million. He also said the team has launched a formal investigation into user complaints about fast quota depletion, even though no anomaly has yet been found, and warned that converting subscription accounts into API traffic for resale or shared access is not supported.

Researchers Jianhao Ma of Tsinghua University and Yuxin Chen of the University of Pennsylvania’s Wharton School have released a paper addressing a roughly 40-year question in optimization theory: how far standard gradient descent can go if its structure is left unchanged and only the step-size schedule is tuned. Their result states that for any pre-specified nonnegative step-size sequence, gradient descent has a lower-bound convergence rate of Ω(T^-1.9319), ruling out the possibility that step-size design alone can match the O(1/T²) rate achieved by Nesterov’s accelerated method. The paper, as described in the report, is notable not only for the theorem itself but also for how the proof was produced. The core argument was generated through iterative work with GPT-5.6 Sol Pro, with the researchers supplying the problem target and a high-level “resisting oracle” strategy, then correcting gaps as they appeared. The result was later translated into Lean 4 code with Codex and formally checked line by line. According to the report, the final formalization used zero “sorry” and zero “admit,” meaning no proof steps were skipped. The code has been made public on GitHub alongside a traceability file linking the paper’s theorems to the Lean implementation.

Gamgee, a startup founded by AI entrepreneur Paul S. Conyngham, said it has raised a $4 million seed round led by Founders Fund to build a repeatable pipeline for personalized mRNA cancer vaccines for dogs. The company grew out of Conyngham’s attempt to design a custom mRNA vaccine for his dog Rosie, who had mast cell cancer and later became the center of a widely shared story on English-language social media. Rosie received an initial shot in December 2025 and a booster the following month, and several tumors reportedly shrank while her quality of life improved. But the remission did not last. After a follow-up surgery, the cancer returned, spread rapidly within four weeks, and Conyngham said he made the decision to euthanize her. Gamgee said the new funding will be used to launch a clinical trial to test the personalized treatment approach first developed for Rosie. The company is working with UNSW, the University of Queensland and the Garvan Institute of Medical Research, with enrollment currently limited to eastern Australia. Conyngham has said the company is also recruiting veterinary oncologists and trial partners. The larger challenge, as described in the source material, is not only whether a personalized vaccine can be designed, but whether sequencing, regulatory review, manufacturing and delivery can be completed fast enough for patients who have limited time.