X / Twitter
Latent Space podcast host / AI Engineer founder swyx
While choosing an insurer for his newly established New Media Lab, the creative studio space housing swyx inc, swyx unexpectedly found that Corgi had an extraordinarily strong reputation among creators. His real estate agent told him, “Corgi now covers every client I have,” approaching 100% of that agent’s new business. He said this rate of penetration was virtually unheard of in an old industry such as insurance. A new insurer focused on the creator economy developing an almost monopolistic reputation represents a classic opportunity: new demand re-segmenting an old market. For peers building independent studios, content companies, or one-person businesses, his implicit advice is not to default to major insurers; small specialists may offer both stronger NPS and a better fit.
https://x.com/swyx/status/2068924451887129055
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux from the OpenAI Codex team revealed that Codex now supports banking weekly/monthly usage resets rather than requiring them to be used immediately as before. He put a product-management-style question to the community: would people hoard them, reluctant to use them, or go all out without worrying? This led to the feedback he really wanted: what parts of the Codex app are not delightful enough and should the team prioritize improving? For heavy Codex users, this is one of the fastest routes to being heard directly by the product team. Anyone with specific frustrations involving reconnecting after interruptions, lost context, CLI behavior, or how billing feels can reply to him directly. From the outside, this pattern of “ship a feature, then publicly gather friction points” is characteristic of how the OpenAI team currently interacts with its community.
https://x.com/thsottiaux/status/2068792010715324444
https://x.com/thsottiaux/status/2068736857312198928
Newsletter author with 150K+ subscribers Peter Yang
Peter Yang reposted an insight from @liu8in on “why HTML is the foundation for agentic video generation.” The author initially wanted to build a video agent but found that LLMs lacked visual intelligence and direct video generation did not work. Switching to HTML suddenly made things work, because HTML is an LLM’s native language: LLMs can express not just information but visual aesthetics through HTML/CSS/JavaScript, with shots, images, assets, and SVG all layered on top of HTML. The accompanying demo generates a product video from any website URL in one click. He also recommended ferryman.io, built by his friend Kevin: write and publish once, and automatically distribute the content to 7+ social platforms. His final post was a personal reflection: the “spend sparingly” mindset he developed growing up as an immigrant has, in the unlimited-token era, become “if I don’t burn through the limit, I’m wasting it.” He has not yet worked out how to handle this mismatch himself. For builders working on video agents, product demos, or automated editing, “LLMs creating visuals indirectly through HTML” is a path worth reconsidering among the options.
https://x.com/petergyang/status/2068755908319236338
https://x.com/petergyang/status/2068854663534031124
https://x.com/petergyang/status/2068874249167884544
Linear Head of Product Nan Yu
Linear Head of Product Nan Yu quoted “Quality is irrational” and elaborated: making a truly high-quality product requires two kinds of “irrationality.” One is irrationally choosing quality every time, even when the short-term ROI does not add up. The other is irrationally believing that controlling everything from top to bottom will produce better results than adopting a generic framework. This is a contrarian bet against the dominant approach of assembling frameworks and copying best practices. Linear’s differentiation in its tool category comes precisely from this top-to-bottom control: animations, keyboard feel, density, and typographic details are all its own decisions. The post thus publicly lays out Linear’s methodology. For peers building tools from 0 to 1, his implicit advice is that frameworks can quickly get a product to a score of 80, but creating something users love “irrationally” requires the irrational choice to build it yourself.
https://x.com/thenanyu/status/2068778750800531640
Vercel CEO Guillermo Rauch
Vercel CEO Guillermo Rauch offered a line worth revisiting: “Coding agents will squeeze every ounce of IKEA effect out of you, if you let them.” In other words, left unchecked, coding agents will extract every last ounce of your IKEA effect. The IKEA effect is the tendency to value something you helped assemble more than its actual worth. He used it to warn that when an agent does most of the work and you change a few lines in the corners, you can still develop a powerful sense that “I made this,” losing objectivity about the result. The implication is that boundaries around which parts you write yourself must be drawn deliberately, or vibe coding can drain both your judgment and your sense of accomplishment. He also mentioned that the v0 team had invested heavily in performance, inspecting and optimizing every frame, from painting and layout to WebGPU shaders and blocking scripts, and promised to fold lessons from the experience back into the Next.js documentation. For heavy vibe-coding users, the first warning deserves more reflection than the second product update: how much of the code you are writing is no longer actually in your head?
https://x.com/rauchg/status/2068838709517336756
https://x.com/rauchg/status/2068778558672273422
Box CEO Aaron Levie
Box CEO Aaron Levie highlighted Sakana’s newly released Fugu model. It exposes a single API externally, while internally a team of specialist models, a mixture of models, collaborates on tasks. It handles simple questions itself and automatically coordinates multiple models for complex ones, with verification and synthesis taking place entirely inside the system. He believes this is the common pattern application-layer AI companies use to build agent harnesses, and that Sakana turning it into an LLM developers can call directly is valuable in itself. As both frontier closed models and OSS models continue to be released, the middle layer that routes tasks to the best model will capture enormous value in the next phase. Levie made another, stronger prediction: agents will use software 100 times more frequently than humans, and one agentic query may retrieve more data than a user would manually access in an entire month. Agents will consume CRM, documents, enterprise knowledge, and analytics data at orders of magnitude greater volumes than people. This creates new requirements: agent guardrails, authoritative data sources, call logs and audits, and human-agent collaboration channels to ensure agents do not miss data or incorrectly alter critical fields. For SaaS companies, this is a signpost: whoever first adapts their product to provide capabilities to headless agents, while adjusting both the business model and technical path, secures a place in the next cycle.
https://x.com/levie/status/2068917230570795178
https://x.com/levie/status/2068851573175021864
Cursor designer Ryo Lu
Cursor designer Ryo Lu added a new Books app to ryOS, his nostalgic browser-based Mac OS. It supports any EPUB file, automatically syncs reading position across devices through the user’s ryOS account, and features a nostalgic wooden bookshelf interface. The development process is revealing: he first built the skeleton in Cursor mobile through mobile vibe coding, then manually adjusted animation and textures until it “felt right.” In other words, the LLM handled 80% of the labor, and the designer injected aesthetics and feel into the final 20%—a firsthand demonstration of a vibe-coding workflow by a product person inside Cursor. For independent product makers, the ryOS series is also a useful case study: one person connecting nostalgic design, practical utilities, and a proprietary account system into a coherent small ecosystem exemplifies today’s “one-person universe.”
https://x.com/ryolu_/status/2068923971136098633
https://x.com/ryolu_/status/2068924375341179347
Y Combinator CEO Garry Tan
Y Combinator CEO Garry Tan returned to his open-source GBrain and offered a judgment for this moment in 2026: looking back, “having your own personal brain and company brain” will prove to have been severely underestimated. His reasoning is that AGI is already on the threshold of usability, and intelligence itself is becoming a commoditized input. The real unlock is not “asking AGI a question,” but systematically capturing private context about yourself and your company; only with this layer of context can AGI create value exclusive to you. In other words, model providers supply intelligence, but you must create the context yourself. This is also his motivation for building and open-sourcing GBrain. For founders and builders, the post amounts to an exhortation: stop leaving conversations only in ChatGPT’s or Claude’s temporary context and seriously build “my own brain” that continuously supplies your context. The earlier you start, the greater the compounding benefit.
https://x.com/garrytan/status/2068701356358308112
https://x.com/garrytan/status/2068701357696323769
Independent builder Zara Zhang
Independent builder Zara Zhang offered a rule of thumb for preventing AI slop: check whether the context you put in is longer than the output. In her own practice, the input is usually 3 to 5 times the output; whenever the input is substantially shorter than the output, the result is almost certain to be slop. The implication is that quality comes from context density rather than prompt engineering. AI does not create value out of nothing; it compresses, reorganizes, and rewrites the material you provide. She emphasized in a follow-up: “I mean context, not prompt.” For builders generating content, designs, documents, or code, this can serve directly as an ROI yardstick: either invest time in building rich context first, or accept that the output will inevitably be average.
https://x.com/zarazhangrui/status/2068923768500793603
https://x.com/zarazhangrui/status/2068964055235321954
OpenClaw co-founder / OpenAI’s Peter Steinberger
OpenClaw co-founder Peter Steinberger responded to community discussion about the project’s current state. Public attention has indeed declined, but internally the team has continued improving quality and expanding, while establishing the project as a nonprofit, in clear contrast to VC-backed competitors with their own commercial agendas. He pointed directly to the data: this was OpenClaw’s strongest week yet. The implicit framing is “we are not driven by hype or fundraising pressure, so we can take our time and get it right.” Two other signals are worth noting: he publicly expressed skepticism about the current popularity of multi-model routing, saying “my intuition was right after all,” and he is trying a new tool for consuming Twitter, effectively moving away from the traditional timeline. For those interested in governance paths for independent AI projects, OpenClaw is offering a new example of “nonprofit + long-term commitment.”
https://x.com/steipete/status/2068961217524490739
https://x.com/steipete/status/2068960117253632160
https://x.com/steipete/status/2068965200343224367
Podcasts
Training Data — Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness
Key takeaway: “What we call a ‘model’ today is no longer the bare weights of 2024, but a whole expanding system. The agent harness looks like the biggest source of alpha in 2026, but within 12 months models will absorb it as a native capability, and the alpha will inevitably move elsewhere.” This is Google AI Studio and Gemini API lead Logan Kilpatrick’s most contrarian assessment of the current LLM industry, and one worth pausing to absorb.
Logan appeared just after Google I/O, the release of Antigravity, and Gemini 3.5 Flash overtaking the previous Pro model through post-training alone. The conversation was densely informative, centered on “how Google competes in the agent era, how models consume the harness, and how a single omni model replaces eight vertical models.”
First, Antigravity is not an IDE but Google’s next-generation general-purpose foundation for agents. Logan explained clearly that Antigravity is simultaneously an IDE, an agent-first web experience, a CLI, and an SDK, while also allowing managed agents to be called through the Gemini API. The same agent harness now runs across Google products including Search, the Gemini app, Cloud, and AI Studio. Previously, Google’s hundreds of products had no common thread. Gemini created a shared “model” thread; Antigravity changes that thread to the “agent harness.” In other words, Google is treating the agent harness as the next generation of Gemini.
Second, “the model eats the harness” is the episode’s central engineering judgment. Logan was direct: “Two years ago, an LLM was a set of weights: tokens in, tokens out. Today we still call it Gemini 3.5 or GPT-something, but it is already an entire system growing around the weights, with agentic tool calling, hosted tools such as search and code execution, containerized sandboxes, and agent harnesses attached around it. Scaffolding moves half a step ahead; then the model consumes it and turns it into a native capability.” The explicit advice for independent developers is not to bet on “writing your own harness” as a long-term source of alpha: models will natively possess these capabilities within 12 months, so that alpha must move elsewhere in advance.
Third, coding is already effectively narrow superintelligence. Logan said plainly, “coding is just so good that it does kind of feel like narrow superintelligence,” and candidly described his experience: “As a developer, I actually have more agency. Ideas that used to be just out of reach are now achievable. The problem now is that I need to set more ambitious goals.” He believes the next fields to see vertical superintelligence will be highly verifiable ones: math, finance, and science. He particularly emphasized science because “making positive impacts happen sooner” is crucial to the broader AI narrative.
Fourth, Omni is a true single model, not routing. Gemini Omni Flash combines capabilities previously covered by 8 independently trained models—text, audio, Lyria music, Nano Banana images, Veo video, and others—into one model: “It does not route to different models; this is a true omni model.” The main publicly available capability at present is video editing. While he was speaking onstage, an audience member used Omni to add a dog jumping onto his lap to the live video. The model even correctly handled the other guest’s subtle expression, looking down at the dog and chuckling. This level of world understanding in the details has left him still “processing what this actually means.”
Fifth, vibe coding has reached significant scale. In one week, 350,000 Android apps were built in AI Studio, and he stressed that “most of those 350,000 apps would never otherwise have been written.” Games are no longer the most popular category: finance and crypto account for 20%, personal productivity for 20%, and generative media for 20%, suggesting the production side of software is opening up.
He closed with a framing worth remembering across the ecosystem: “There's never been more opportunity to go and build something. At the same time, the models are doing more than they've ever done before.” Focus remains a startup superpower: large companies are forced to do many things and therefore cannot focus, which is precisely the opening for vertical builders.