X / Twitter
Swyx, AI Engineer Conference Organizer and Latent Space Host
Swyx posted a series of previews from the AI Engineer conference, inviting attendees to AIE Expo at 12:30 for a “flash surprise,” followed by a live Latent Space podcast with chip company Etched on Expo Stage 2. He also shared an observation: the most enthusiastic applause during this year’s AIE keynotes was for a discussion of men openly addressing emotions and mental health in hypergrowth environments. It came from his live Q&A with Mike Krieger around Fable and Tag. He was proud to see the topic normalized at a technology conference. For those following AI engineering culture, this signals a community beginning to confront hypergrowth’s psychological costs.
https://x.com/swyx/status/2072760421627597198
https://x.com/swyx/status/2072754722059239471
Boris Cherny, Anthropic Claude Code Team
Boris called Artifacts in Claude Code “life-changing” and announced that the feature would soon expand to Pro and Max subscriptions. Artifacts present Claude Code output as standalone interactive experiences rather than terminal text streams alone. A core team member actively promoting it suggests Anthropic is extending Claude Code beyond a purely developer audience to broader subscribers. Pro and Max users should watch for availability.
https://x.com/bcherny/status/2072777472970563995
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault offered a short but substantive teaser: GPT-5.6 Sol Ultra is coming. His words were: “Can’t wait to see what you do with it. Save your hardest prompts.” This suggests a substantial improvement on difficult reasoning problems, with an official team member inviting users to prepare their toughest challenges. Given his responsibility for both Codex and ChatGPT, the release will likely reach coding and conversational products. Competition between OpenAI and Anthropic’s high-end models continues accelerating.
https://x.com/thsottiaux/status/2072607914217320644
Peter Yang, AI Tutorial Creator
Peter shared three ways to maximize Fable’s value before July 7: prepare context with cheaper models, use Fable for planning and other models for execution, and lower reasoning effort to Medium while monitoring its work. He also released a full tutorial with 5 “Fable-worthy” use cases. Another post summarized audience voting for a Nous Research interview: the top five topics were Hermes’s origin story, agent memory and persistence, new real-world use cases, team operations and delegation, and long-term vision and differentiation. The episode will arrive in the next few weeks. The most interesting third post described a Codex project with his 8-year-old daughter. She drew a dragon, uploaded it to generate variations in different poses, gave Codex voice feedback throughout, and sent the results to a custom-sticker site for printing—about $20 for 10 stickers. This offers a ready-made template for turning AI into a family activity during summer holidays.
https://x.com/petergyang/status/2072842766053499353
https://x.com/petergyang/status/2072838004310507975
https://x.com/petergyang/status/2072756657856422379
Cat Wu, Anthropic Claude Code and Cowork Team
Cat revealed Claude Tag’s internal adoption at Anthropic: engineering, product, data, sales, and marketing all use it, and the internal version already accounts for 65% of product PRs. She and a colleague recorded a conversation covering the CEO/CTO rollout playbook, why security was designed in from day one, and implications for future work. The commercial offer is aggressive too: Claude Enterprise organizations can claim $25000 in credits and Claude Team organizations $2500 to try Claude Tag before September 1. The 65% figure is among the most aggressive public figures for AI-written code at major technology companies. Leaders introducing agent workflows have a low-cost opportunity to experiment.
https://x.com/_catwu/status/2072743070316257662
https://x.com/_catwu/status/2072731500928508331
Thariq, Anthropic Claude Code Team
Thariq addressed widespread questions about Fable subscription availability. The clear timeline is that Fable will be removed from subscriptions after July 7. He stressed that this is not permanent: the team aims to restore it as a standard subscription component as soon as compute capacity allows, as promised in the original blog. The removal therefore reflects capacity rather than a strategic reversal. Subscribers heavily reliant on Fable should arrange a transition before July 7 while anticipating its eventual return.
https://x.com/trq212/status/2072814903170408784
https://x.com/trq212/status/2072814904210509905
Guillermo Rauch, Vercel CEO
Guillermo compared AI Gateway to a “Token Delivery Network,” like a CDN for models. He described a practical problem: Fable’s sudden retirement raised concerns about production traffic breaking, and retirements will become more frequent amid GPU shortages. Data shows healthy production traffic still running on many old versions. The answer is the newly released AI Gateway Rules: dynamically rewrite model routes without redeploying, for example redirecting anthropic/claude-fable-5 requests to anthropic/claude-opus-5 with one command. His business logic is direct: losing tokens means losing revenue and customers. Another post introduced private connections between services, registering bindings in vercel.json with one line and reading internal URLs through environment variables, with Node, Python, and Dockerfile support. Hot-swapping model routes is essential for teams running multiple-model applications in production.
https://x.com/rauchg/status/2072741369848746315
https://x.com/rauchg/status/2072715658157027375
Aaron Levie, Box CEO
Aaron wrote at length about enterprise AI implementation, arguing that deployments beyond chatbots unquestionably require substantial work aligning systems with underlying business processes. Existing workflows were never designed for agents: data is fragmented, legacy systems cannot connect, and institutional knowledge remains undocumented in employees’ heads. Reliable deployment at scale requires cleaning data, modernizing IT, building evals, managing change, deciding where humans remain in the loop, and identifying the company’s new IP. His conclusion points to a career trend: this is why applied AI companies are expanding FDE teams—forward-deployed engineers—and dedicated deployment firms are forming. FDE will become one of technology’s most important roles. For engineers seeking enterprise AI opportunities, this is a clear career signal.
https://x.com/levie/status/2072875685811716182
Matt Turck, FirstMark Capital Investor
Matt released an in-depth conversation with NVIDIA’s Bryan Catanzaro about Nemotron and NVIDIA’s AI lab. The central question is pointed: why does a chip company employ hundreds of researchers to build models and give them away? Topics include whether open AI is catching the frontier, whether the US trails China, why enterprises actually choose open models, Catanzaro’s then-radical 2008 bet on machine learning on GPUs, and working with Andrew Ng and Dario Amodei at Baidu. Technical topics include training a 550B-parameter model with 4 bits, an accessible explanation of Hybrid Mamba-Transformer architecture, predicting 5 tokens at once through multi-token prediction, and why NVL72 was designed around mixture of experts. They also discuss Nano, Super, and Ultra positioning, NVIDIA’s bet on speed for agents, and the importance of a 1-million-token context window. For readers seeking NVIDIA’s model strategy and frontier training techniques, this is a substantive episode available on Spotify, Apple Podcasts, and YouTube.
https://x.com/mattturck/status/2072723410975629364
https://x.com/mattturck/status/2072723415870411232
Zara Zhang, Builder
Zara posted three sharp observations about AI and learning. First: AI slop comes from having nothing substantive to say, not bad style. Second, a recent graduate described feeding course materials to AI and learning from it rather than attending human lectures, often finding AI taught better than professors. Third, one of the best things users can do for agents is let them converse in groups instead of private one-to-one chats. The consistent theme is that AI’s value depends on the substance supplied and how it is organized. Group-chat collaboration is worth testing in multi-agent workflows.
https://x.com/zarazhangrui/status/2072943922385715262
https://x.com/zarazhangrui/status/2072729444943577601
https://x.com/zarazhangrui/status/2072726336158998760
Nikunj Kothari, FPV Ventures Partner
Nikunj wrote to “AGI summer tourists.” As last year, many people briefly visit San Francisco and leave calling it soulless, nothing but 996 culture and AI, dirty, chaotic, and troubled by homelessness. He acknowledged some truth in all these criticisms, but argued against judging a city after days or even weeks. His advice: stop being a tourist and spend real time living there. Booms and busts happen, but it can be wonderful; if it truly does not suit you, go home and build your hometown into what you want. He signed as “someone who has lived here for 14 years.” It is an insider’s perspective on commentary around SF’s AI boom.
https://x.com/nikunj/status/2072780155924480074
Dan Shipper, Every CEO
Dan voiced a common experience among heavy Fable users: it can run alone for hours, then return only two paragraphs explaining what it did. He endorsed the view that AI needs better ways to “tell stories” to humans. This exposes a real product gap: autonomous work duration is growing faster than observability and narrative capability. People struggle to trust colleagues who give conclusions without explaining the process, and agents are similar. For agent builders, this identifies an unresolved need: interfaces that narrate progress through long-running tasks.
https://x.com/danshipper/status/2072805884376301737
Claude, Anthropic’s Official Account
Anthropic announced two things. First, a video conversation with Boris Cherny and Cat Wu covers the evolution from Claude Code to Claude Tag and its spread from engineering across Anthropic, alongside confirmation that Claude Fable 5 is available in Claude Tag. Second, Built with Claude: Life Sciences is a global online hackathon with Gladstone Institute, focused on life-sciences research and development using Claude Science and Claude Code, with $100,000 in credits as prizes. It runs July 7 through 13, with registration closing this Sunday. Teams working in biomedicine and familiar with AI tools have an accessible opportunity.
https://x.com/claudeai/status/2072725610061803522
https://x.com/claudeai/status/2072681853971001849
https://x.com/claudeai/status/2072681856730792282
Podcasts
No Priors — How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor
Key takeaway: Nuclear energy is a hardware-execution problem, not a design problem. Whoever can manufacture reactors at scale and iterate rapidly like Toyota Camrys can reduce energy prices tenfold.
Isaiah Taylor is Valar Atomics’ founder and CEO; his great-grandfather was a nuclear physicist on the Manhattan Project. The company, less than three years old, has just achieved in Utah something previously accomplished only by national laboratories and century-old corporations: the first electricity-generating advanced reactor built by a startup, the first US TRISO reactor started in over 50 years, and the fifth new US nuclear installation generating power since 2000. With AI compute driving electricity demand dramatically, the company explains why nuclear revival has never been constrained by physics so much as the ability to build things.
The first key is a regulatory breakthrough. For twenty years, nuclear energy faced a loop: permits required operating data, but operation required permits. The industry responded by becoming a “modeling and simulation industry” producing precise paper reactors. Valar took another route: US law has long provided two paths. NRC oversees commercial deployment, while the Department of Energy’s predecessor ERDA was established to test reactors—a testing path unused for forty years. Executive order EO14301 required three advanced reactors on US soil to reach criticality before July 4. Valar started under DOE authorization and now runs at 100 kilowatts, splitting approximately 10 to the 17th power atoms per second.
The second key is a shift in safety philosophy: traditional nuclear power controls risk by reducing accident probability; Valar aims to eliminate consequences. Its reactor uses helium cooling, graphite moderation, and TRISO fuel. Its regulatory safety baseline assumes everything at the plant fails, yet the public and workers receive no radiation dose. A few days after recording, they planned to demonstrate shutting down all plant power and safety systems after an emergency stop, with a water jacket passively removing decay heat through natural circulation and cooling the reactor over two days without moving parts or human intervention. Taylor also offered a counterintuitive statistic: by deaths per unit of electricity, nuclear is safer than solar because more people die falling from roofs while installing panels.
The third key is quantifying speed. The internal metric is tick rate: 2 years and 4 months from incorporation to first atom fission, then only 7 months to the second, with a goal of starting a new reactor every few minutes. Engineering innovations include Modular Citadel prefabricated concrete biological shielding, with sinusoidal joints preventing radiation paths, requiring neither grout nor bolts. Work traditionally taking 3 months was stacked in 42 hours. The two engineers who invented this radiation-shielding concrete without rebar were only 23 and 21. The starkest story concerns reactor protection: a supplier quoted $5 million and two and a half years; Valar put 5 engineers in a conference room for 6 weeks and built it for about $400,000. The supplier then spread claims that Valar was unsafe and would kill people. Taylor’s assessment was unsparing: “The deeper you look into the nuclear industry, the more you find this everywhere. It’s a fake industry that hasn’t built anything in forty years, and the few products it has still command a 100-fold premium.”
Commercially, Valar rejected the traditional startup path of assembling paper plans to seek project financing. It uses risk-tolerant equity to build reactors on its own balance sheet, first creating cheap power at its own gigasite and attracting data-center workloads. It also staged a marketing demonstration: directly powering an NVIDIA Blackwell chip with the reactor and hosting nuclearwebsite.com on it. The page reports how many uranium atoms were split to transmit it; when the reactor stops, the site disappears. Valar merchandise is sold only there.
Taylor’s ultimate judgment is memorable: AI and robots are converting manufacturing labor into energy consumption. Following the supply chain upward, energy is the sole fundamental input. Every tenfold reduction in its price brings us closer to hyper techno industrialism where “almost everything is free.”