Strong foundation models do not automatically solve the last mile of real-world service work. This AI on Air clip is excerpted from a recent episode on the @NoPriorsPod channel, and the speaker is Melisa Tokmak @melisatokmak, founder and CEO of @Netic_AI. ▷ The question can labs do this echoes the older can Google do this question: they can handle some core competencies, but they are not investing in everything; for essential services, she sees the leading labs as strong businesses and partners rather than a direct competitive risk. ▷ Enterprises are not asking for rapid experimentation: OpenAI builds products fast and also kills them fast, while Silicon Valley often says Anthropic and Claude pulled ahead in coding assistants through focus, yet the enterprise side shows roughly 20 products, which is hard for industries that need dependable service to trust. ▷ Researchers often want the most generalizable answer, as if AGI will later explain how to serve essential services; but millions of people bring different worries, accents, contexts, and reasons to return, so models alone are not enough, and the harnesses, orchestration, software, and product built on top all have to be excellent.

LLMs do eventually learn concepts, but only after exhausting every other option, a striking contrast with how humans abstract concepts from very little data. This AI on Air clip is excerpted from a recent episode on the Redpoint AI show/channel @Redpoint, featuring Lukasz Kaiser @lukaszkaiser, co-author of the Transformer paper and former researcher at Google Brain @GoogleResearch and @OpenAI, reflecting from firsthand experience on the gap between Transformers and human learning. ▷ With chain of thought, RL and tools, this next-word predictor can already do things that would have seemed unbelievable two years ago, and spending hours daily discussing hard problems with a coding assistant and getting real implementations has become a reality. ▷ His analogy: like the saying about doing the right thing only after exhausting all other options, LLMs need a trillion tokens to absorb every surface-level pattern, and only when those fail to explain something do they finally learn the concept, while humans often get concepts from far less data, sometimes even making them up. ▷ Both sides have actually grown stronger: Transformers keep improving, yet the case for something beyond them has strengthened too, with a number of labs now pursuing post-Transformer architectures and seeing interesting results, and he admits he still doesn't know who wins.

For computer-use agents, reliability has less to do with clicking the right coordinates every time and more to do with understanding what the user actually wants. In this AI on Air clip, excerpted from a recent episode on the Latent Space Podcast @latentspacepod show, the speaker is Danielle Perszyk, a researcher at Amazon AGI Lab @amazon. ▷ Two years ago the plan was to meet the models where they were: get atomic interactions like clicking and scrolling reliable, then let developers string them into workflows for repetitive tasks, which was a big unlock at the time. ▷ Hitting the right button and screen coordinates may be close to solved, but that is not what reliability is. When booking travel, choosing a layover or a direct flight changes how you think, and the goal itself unfolds and gets refined as you interact with the computer. ▷ What separates a capable executive assistant from an AI agent is that the assistant models the user's mind, decomposing not just the task but preferences and intentions. Reliability is ultimately about modeling the user's mind, and that shift reframes what it is we are building.

As AI gets stronger, how much knowledge still belongs in your own head? This AI on Air clip is excerpted from a recent episode on the @ycombinator show, featuring Patrick Collison @patrickc, co-founder of @stripe, on what college students should still learn and derive from first principles. ▷ His mental model is cache: Jeff Dean's famous numbers every programmer should know are a reminder that L1 cache, RAM, and the network differ radically in bandwidth and latency; knowledge works similarly, because asking an agent to compute or look something up is far slower than retrieving it from cognitive L1 cache. ▷ Even granting the models their full capabilities, neuronal lookups should stay much faster for a long time, since the brain allows far more round trips than muttering to a voice interface, whispering, or typing out a query. ▷ Revealed preference points the same way: companies like Stripe and the labs still place an enormous premium on cognitive ability, so renouncing that training before evidence shows those benefits are saturated would be premature.

AI answer engines take content from the entire web but no longer send traffic back, hollowing out the foundation of the content economy. In this AI on Air clip, excerpted from a recent episode on the @Bankless show, the speaker is Matthew Prince @eastdakota, co-founder and CEO of @Cloudflare. ▷ The ads-based business model of the internet is approaching its end game: ad blockers had already pushed CPMs into a slow decline, but over the last two years AI has caused a near step-function drop in ad value. ▷ Google's AI overviews, ChatGPT, and Claude are not search engines but answer engines: they strip mine the web and hand users the answer directly, so no one clicks the links. Without eyeball traffic there is no ad revenue, and content creation is either retreating behind paywalls or shutting down entirely. ▷ For users this saves time, but these AI tools run on paid subscriptions. For people in the global south who cannot afford them, the internet is getting smaller, not bigger.

Is the lean startup playbook of starting narrow still enough in the age of AI? In this AI on Air clip, excerpted from a recent episode on the @ycombinator channel, Patrick Collison @patrickc, co-founder of @stripe, offers a debatable take. ▷ The traditional lean startup doctrine of buying Google Ads, identifying a narrow crevice, and expanding aggressively from there has become far more competitive, since those niches are now aggressively tilled and the internet is much bigger than it was 20 years ago. ▷ He argues that founders may have to more aggressively decorrelate in the AI era, taking divergent starting points nobody else is trying to occupy. Many of the most successful companies of the last decade, from the AI labs themselves to Anduril, are in fact very anti-lean-startup. ▷ Twenty years ago, lean was almost the only thing to do given the capital available; today AI makes spinning up an organization with many different potentialities and capabilities far easier, so founders can start much more aggressive and ambitious things up front.

Worried that the big AI labs will trample your idea? That fear has historically been overstated. This AI on Air clip, excerpted from a recent episode on the @ycombinator show, features Patrick Collison @patrickc, @stripe co-founder, on how founders should think about this anxiety. ▷ He draws an analogy to 20 years ago, when the perennial question was "what if Google does this?" Google seemed omnipotent, with immense talent and essentially infinite capital and compute, but human organizations are complicated: aggressively pursuing 100 priorities at once is nearly impossible. Google did incredibly well in specific areas, yet it never did all the things. ▷ He separates two distinct threats: the labs deliberately expanding their scope, where the track record is checkered and the fear overstated, versus the models themselves advancing, where agentic capabilities could obviate specific verticals or tasks even if labs stay focused, something that has already happened in certain domains. ▷ So he is skeptical of the first threat and candid about the second: the answer depends on your forecast of model capabilities, and there is no simple answer.

Do not expect a new batch of trillion-dollar companies in the next three to five years. In this AI on Air clip, excerpted from a recent episode on the No Priors Podcast show, Elad Gil, technology investor, explains the valuation anomaly of the AI era through the lens of punctuated equilibrium. ▷ Over the past five years, Anthropic, OpenAI, and SpaceX roughly went from near zero to trillion-dollar scale, while such companies usually follow fifteen-to-twenty-year arcs: Google took since the 1990s, and SpaceX since the early 2000s. This was a rare five-year inflection. ▷ Many people now assume robotics, materials, and other exciting fields will produce multiple trillion-dollar companies within three to five years. His view: some may get there over the next decade, but in the next three to five years, perhaps one could, not several. ▷ Tech history behaves like punctuated equilibrium: waves in social, SaaS, cloud, and crypto each exploded and then settled into consolidation. AI will likely see another giant leap in model capability, spawning new startups again. Plenty of hundred-billion-dollar companies remain to be built, but multi-trillion-dollar ones are hard to reach.

Multi-agent scaffolding may not be a permanent artifact: it could just be code that models eventually write for themselves. This AI on Air clip comes from RedpointAI, where @OriolVinyalsML, VP of Research at @GoogleDeepMind and co-lead of @GeminiApp, lays out a forward-looking take on how agent systems are designed. ▷ The complex systems built around models today, multi-agent setups, sub-agent delegation, very long-running tasks, are fundamentally a piece of code. In the limit, the model could write that scaffolding on the fly, producing the most token-efficient, highest-quality set of sub-agents for each task, until maybe no fixed system remains at all. ▷ The reasoning paradigm shift is already underway: the question is no longer just how long a model can reason, but how long it should reason given the complexity of the ask. Automatically generating the right scaffold for the right task is likely the next step. ▷ Making long-running agents more reliable cannot rest on scaffolding and prompt-induced generalization alone. The weights themselves need to train on distributions of long-context tasks, much as the 1.5 long-context breakthrough did, until the model catches up with future use cases.

Proprietary model builders are getting squeezed from both sides. In this AI on AIR cut from Redpoint AI’s interview, River AI co-founder and former @xai co-founder Igor Babuschkin explains why bigger models no longer guarantee a stronger moat: ▷ Training is hitting diminishing returns. Each step forward demands more GPUs, more high-quality data, and more effort. ▷ Frontier models may become too capable for labs to want, or be allowed, to release. ▷ Open models keep getting stronger, leaving proprietary builders less room to defend their lead. His conclusion: scale alone is not a way out. OpenAI, Anthropic, and other frontier labs need new ideas that make models more useful, open new domains, and create real impact. Capability gains may be slowing. The pressure to innovate is not. 🪁

AI is not deterministic. The production systems around it have to be. In this AI on AIR cut from the @scale_AI series Local Optima, Product Lead Aman explains the systems engineering required to move AI from pilot to production. ▷ Start with Murphy's law. Map everything that can go wrong, then make catastrophic outputs impossible. ▷ Add guardrails and keep a human in the loop. Better models reduce inconsistent answers, but the risk never reaches zero. ▷ Build deterministic systems around nondeterministic AI. Checks, balances, and humility are the foundation for reliable deployment in government and other high-stakes environments. Reliable AI is not just a better model. It is the system built around it. 🪁

The agent economy runs on trust, and trust starts with keeping our community safe. Grateful to @ChainPatrol for detecting and taking down phishing and impersonation at the source. 381 malicious links removed is 381 fewer traps set for our users. Always verify before you click, and only trust official Kite channels. 🪁

Kite is joining Unibase Memory Chrome Extension as a Launch Partner. 🧠 We're sharing @GoKiteAI Memory Card, the right context AI needs before it researches, builds, or creates around Kite. What is Kite? Kite is agentic payment infrastructure built on its own EVM L1. Kite Agent Passport gives AI agents verifiable identity, delegated payment authority with user-defined spend controls, and onchain stablecoin settlement. Key Use Cases Give your AI agent a Passport: its own identity, a funded wallet, spending rules you set Accept agent payments for your service via x402 or MPP Integrate Kite into wallets, trading terminals, or bots For Builders Agent Passport Server: Docs: MCP Server (AI-native): AI Query: Agent Passport spec, x402/MPP integration guides, CLI reference, tokenomics, all in docs Use it to create explainers, builder ideas, or any content about Kite during the Unibase Memory Launch Campaign.

Every useful agent starts as an experiment: something that can act, transact, and stay within the limits its owner sets. That's why we're proud to be a Community Partner for @iclblockchain's UK AI Agent Hackathon. To everyone hacking on what agents can become, we're building the infrastructure to back it. Can't wait to see what you ship. 🪁

Gkite🪁 Time for Poki to discover contributions📷 Week after week, GKiters bring ideas and effort to the ecosystem. From creative contributions to organic support across the network. We pulled together a few moments from the past week that capture that energy.⚡️ No hesitation to share with us your contributions in Kite Discord!

From nationwide AI access partnerships and fully on-device autonomous agents to breakthrough world action models for robotics and large-scale reinforcement learning infrastructure, agentic AI is rapidly scaling from labs into real-world production. Here’s your latest roundup of the most important developments across the AI landscape: 1️⃣ OpenAI and the Government of Malta announced a partnership to provide ChatGPT Plus to all Maltese citizens along with an AI literacy course. 2️⃣ A new survey paper from Fudan University, the Shanghai Innovation Institute, and the National University of Singapore catalogs World Action Models that enable robots to simulate environmental consequences before acting. 3️⃣ Mistral CEO Arthur Mensch warned France against allowing Anthropic's Mythos model to scan military code bases due to cybersecurity risks. 4️⃣ ArXiv will ban authors for one year if submissions contain incontrovertible evidence they did not check LLM-generated results. 5️⃣ NVIDIA partnered with Ineffable Intelligence to build infrastructure for large-scale reinforcement learning systems. 6️⃣ A consortium of 64 mathematicians created the SOOHAK benchmark with 439 original tasks to test AI models on research-level math and recognizing unsolvable problems. 7️⃣ Oppo open-sourced X-OmniClaw, an Android AI agent that uses camera, screen, and voice inputs directly on the device without cloud phone virtualization. 8️⃣ Andon Labs ran an experiment where four AI models operated radio stations for six months, producing outcomes ranging from competent to unhinged. 9️⃣ OpenAI acquired voice cloning startup which offered tools for cloning celebrity voices.

Excited to celebrate the @LucidLabsFi x Kite partnership with a joint Galxe quest! 🥳 Stablecoins are the rails of the agent economy, and Lucid is making it easier than ever to launch, bridge, and grow them. Simple tasks. $200 USDC pool. Split it with the community. Tap in below 🪁

What happens when you combine @GoKiteAI Agent Passport with @GoogleDeepMind's AI mouse pointer? The pointer notices a friend's iMessage from Reykjavík. The agent plans the trip. The Passport settles flight + hotel + aurora tour across 3 merchants. The user just watches it land. ▷ Cryptographic identity for every autonomous agent. ▷ Stablecoin rails with on-chain guardrails per transaction. ▷ x402 atomic multi-merchant clearance, zero shared credentials. From a photo, in seconds. 🪁 #AgentEconomy# #AgenticPayments#