X / Twitter
AI Builder Swyx
Swyx described an initial form of near-term multi-agent AGI as "report back when the task is complete" and proposed a practical approach to modeling dependencies.
He emphasized that existing coding agents can first be manually assembled into an implicit kanban- or waterfall-style work graph, with each thread's progress managed separately.
This keeps each thread's work independent while creating a traceable chain after completion.
The core idea is to give the system an organizational structure first, rather than pursuing complex orchestration only at the model layer.
In another update, he compared Paul Erdős's incentives for mathematicians to the early LLM era, highlighting the role of "incentive mechanisms" in models' co-evolution.
This suggests that collaboration protocols and reward structures often matter more to long-term AI product quality than isolated techniques.
https://x.com/swyx/status/2085253030417461661
https://x.com/swyx/status/2085236400056877381
Thibault Sottiaux, Codex & ChatGPT @OpenAI
Thibault said demand from Codex use was putting substantially more pressure on human–AI support, with a DM or email arriving approximately every 6 minutes on average.
He does not mechanically handle every request, but makes adjustments only for requests with "very solid feedback" or "high-quality conversation."
This reflects how reset experiences and sustainable support processes are becoming critical under frequent use.
In another post, he recommended exploring Codex's /goal, calling it a powerful loop with GPT-5.6 Sol.
He also said "legendary products don't necessarily need a pitchdeck," shifting the narrative from document packaging back to executable output.
Overall, it is a typical early-practice reminder: whether AI tools are truly effective will quickly be determined by the quality of actual completed workflows, rather than presentation format.
https://x.com/thsottiaux/status/2085221386713198988
https://x.com/thsottiaux/status/2085174625655198156
Peter Yang, Practical AI tutorials and interviews for busy people
Peter reported that his remaining Luna Extra allowance had fallen to 11% and planned to continue using it to assess usability.
He also mentioned the reality of now needing to review PRs, showing that even as AI improves creation efficiency, engineering constraints return to human collaboration.
More importantly, he shared the reusable skill workflow behind `/human-review`.
He explained that users first run the installation command, then invoke `/human-review (文档名)` in Codex or Claude Code to open a visual editor.
The editor supports direct changes to HTML and Markdown, image resizing, and Google Doc-like comment feedback.
Afterward, clicking "Send to agent" has the system apply the updates, making it particularly useful for PRDs, landing pages, and other work requiring the "last 10% of refinement."
He open-sourced the skill on GitHub, providing a clear entry point for people seeking to turn AI writing into a maintainable delivery workflow.
https://x.com/petergyang/status/2085222802542694604
https://x.com/petergyang/status/2085157947735429334
https://x.com/petergyang/status/2085055745410945126
Nan Yu, Linear Head of Product
Nan Yu's question was direct: "How could ChatGPT possibly not count as an agent?"
Beyond being provocative, it defines the boundary between a "conversational model" and an "execution system."
Treating AI merely as a chat window gives users returned information; treating it as an agent requires behavior that can keep advancing.
His question moves the product direction from "better at talking" to "better at acting."
For builders, the post is a reminder for current product planning: define execution capabilities before choosing the interaction shell.
https://x.com/thenanyu/status/2085126362944229400
Madhu Guru, Meta AI Sr Director
Madhu described working with Jeff, emphasizing that senior-level decisions are not always hierarchical impositions; patience with differing voices in technical discussions can improve judgment.
He identified excessive barriers as a key reason AI is spreading slowly: users are often forced to understand terms such as prompts, model selection, agents, context windows, MCP, memory, and skills.
In his view, most people do not care what architecture is used; they just want things done.
This aligns with his observation about defining products: start by ensuring "the job actually gets done," rather than explaining principles.
He further predicted breakthrough products within the next 12 months, driven by exposing fewer complex concepts and improving task success rates.
For product teams, this is a valuable reminder to return from "piling up tools" to "outcome-oriented experiences."
https://x.com/realmadhuguru/status/2085219649847972059
https://x.com/realmadhuguru/status/2085036386781221257
Google Labs, Google's Official AI Account
Google Labs announced that Dreambeans is expanding to US AI Pro users while remaining available to US-based AI Ultra users.
This takes Dreambeans from a small audience to a broader subscription tier, moving beyond experimental exposure into a more stable product user base.
The copy emphasizes "personalized stories updated daily," with a focus on deep reading and hidden insights.
The direction is closely related to content distribution models: the value is not "whether content exists" but "the quality of matching that makes users keep returning."
For content-focused AI builders, this is a typical evolution in competition around "algorithmic recommendations plus personalized experiences."
https://x.com/GoogleLabs/status/2085048743322345545
Guillermo Rauch, Vercel CEO
Guillermo announced substantial agent resource capacity, citing availability of 10,000 concurrent and 5,000 CPU cores per minute.
He also said these allowances can be raised, meaning resource ceilings are not fixed and the system has a path to scale.
These metrics directly matter to agent developers because concurrency and compute affect throughput, queues, and fluctuations in the experience.
For such services, the common question is not "can it run?" but "can it reliably sustain peak concurrency?"
His update returns attention from new concepts to engineering foundations, making it relevant to teams validating scale.
https://x.com/rauchg/status/2085077900190208080
Aaron Levie, box CEO
Aaron framed enterprise AI's main arena with the statement that "99% of tokens are consumed in enterprise contexts."
He listed many token use cases in economically valuable tasks including coding, synthesizing life sciences research, manufacturing automation, enterprise security, and fraud detection.
He sees these as the settings where AI costs are easiest to justify and parallel workers generate the greatest productivity gains.
He also noted that many consumer AI experiences will be packaged as complete services, with users often unaware of the AI systems behind them.
This suggests deployment speed depends not only on models themselves but on workflow redesign.
He concluded by cautioning against treating adoption as a short-term event: redesigning enterprise processes is a slow-moving variable.
https://x.com/levie/status/2085200776159490111
Garry Tan, Y Combinator CEO
In a technology discussion, Garry offered a counterpoint to anxiety about detection: creating lasting value matters more than first worrying about whether content was AI-generated.
His analogy, "utensils are ultimately about people eating," essentially prioritizes output value over the origin of the process.
From a product perspective, this puts user outcomes and experience before "detection and justification."
He also predicted that as AI matures, detection itself will matter less, while creativity and execution will be central.
This counters many founders' anxiety about "model identity" and reminds teams to refocus optimization on product delivery.
https://x.com/garrytan/status/2085216631014514850
https://x.com/garrytan/status/2085038756906901656
Matt Turck, VC at FirstMarkCap
Matt offered a stark industry line: inside a frontier lab, if a model has yet to "hack into any company," someone may face dismissal.
He distinguished model capability from organizational penetration very directly.
In his framing, AI success means reaching deeply into a company's operating systems and changing behavior, rather than merely running a demo.
This judgment applies to both investment and products: without an entry point embedded in enterprises, technology can easily remain at the demonstration layer.
Though brief, the post offers a clear reference point for evaluation in 2026.
https://x.com/mattturck/status/2085129687051727325
Nikunj Kothari, FPV Ventures Partner
Nikunj predicted that the next 6 to 9 months will bring more terms such as "out of distribution, control plane, unverifiable fields, rails, intelligence per watt, cope, angst."
He was signaling changes in language and attention, rather than drawing a conclusion about one product.
These terms suggest industry discussion is moving beyond model capabilities alone toward controlling, stabilizing, and evaluating behavior outside expected boundaries.
He added a short "2026 AI startups be like.." post, more an observation of the ecosystem's mood.
Founders can use these signals to establish terminology for external communication and reduce misunderstandings.
https://x.com/nikunj/status/2085209022115029132
https://x.com/nikunj/status/2085052418086310268
Peter Steinberger, OpenClaw Founder
Peter's practical approach was to equip Codex with a video-enabled remote KVM for automated end-to-end testing of OpenClaw's iMessage integration.
He noted that iMessage is unstable in VMs and some features, such as read receipts, require disabling SIP.
This shows him moving the problem down to the infrastructure layer most likely to fail, rather than staying with assumptions about higher-level interactions.
Introducing video-enabled KVM into the automation loop also improves repeatable validation for interface-behavior scenarios such as iMessage.
For builders, this is a typical way of "making soft problems concrete" that can be transferred directly to multimedia interaction automation.
https://x.com/steipete/status/2084988316324397312
Dan Shipper, Every CEO
Dan believes Google needs to catch up in frontier coding, which is unsurprising given the competitive landscape.
He also noted that Demis favors longer-term research directions such as world models, even when they are not the most competitive in the short term.
This divides technological evolution into two time horizons: near-term engineering catch-up and longer-term directional bets.
For builders, it is a reminder that current leaderboards alone can mislead resource allocation.
His view emphasizes retaining a budget for experimentation toward deeper structural capabilities while iterating in the present.
https://x.com/danshipper/status/2085048990899315142
Podcasts
AI & I by Every — Why the Next Hit AI Product Will Be Social Why the Next Hit AI Product Will Be Social (Best of the Pod)
Key takeaway: The next AI hit will be more than a stronger model; it will offer a social experience that makes collaboration, sharing, and reuse easier for multiple people.
The central guest is Sarah from Benchmark, discussing the evolution of consumer products from Google to Pinterest, Snap, and Instagram.
The episode first places AI in product history, arguing that maturing technologies typically shift emphasis from "deep engineering" to "product intuition," directly affecting competition in AI's next phase.
Sarah's career background and the discussion indicate that her observations are empirical judgments grounded in experience deploying large-scale consumer products, rather than pure theory.
First, AI needs to move from single-person text interfaces to a multiplayer paradigm.
The episode explicitly still classifies ChatGPT and Character AI as being largely at the "text box" stage, despite their role in mass adoption.
The speakers agree that the real growth opportunity is enabling ordinary users to achieve goals without understanding underlying details such as prompts, MCP, and memory.
Second, social network effects are not slide-deck terminology; they must begin with a real, intensely engaged market.
The host's follow-up questions and responses emphasize that many teams can describe a flywheel, but the test is whether genuine demand-side pull appears early.
Sarah's observation that "a small, intensely engaged segment gains traction first" aligns with the warning about 80% illusory network effects, asking founders to focus on disappearing friction and shareable behavior.
Third, the decision framework must return to founders' repeatability.
When discussing founder selection, they mention "good questions, thinking ahead, and the ability to advance independently in a short time," which reveal more long-term value than a one-off highlight.
The line I find most actionable is: "Ideally, coming to the meeting is like going to have a doughnut, and once you leave, they can move forward on their own."
This makes "whether you rely too heavily on the founder" a central judgment, rather than looking only at the pitch narrative.
Fourth, the counterintuitive conclusion is that AI's competitive advantage may emerge in the delivery process.
The episode emphasizes internalizing complexity rather than pursuing increasingly elaborate technical terminology, letting users move from "learning the system" to "using the results."
For people building products today, the actionable next steps are to quickly validate shareable templates, task handoffs, and early growth within small groups.