AI agents are reshaping industries. How can students prepare for the opportunities ahead? Our APAC Lead @0xLaughing will join students and industry speakers at “AI Unlock: Students in the Agent Era” to discuss AI, Web3, and career development. Hosted by @tbablockchain, @trondao, NTU Fintech Club, and @GoKiteAI. Speaker lineup: ▷ Shih-Wei Liao | Associate Professor, Dept. of CSIE & GINM, NTU ▷ @gobananas929 | Operations Manager, TRON ▷ @Alvin0617 | Founder, CryptoWesearch ▷ @0xfomor | Social Lead, TrueNorth AI ▷ @0xLaughing | APAC Lead, Kite September 8, 18:30 to 21:30 (UTC+8) Room 102, 1F, Xinsheng Lecture Building, National Taiwan University Join us to explore opportunities in the agent era. 🪁

Trusted agent payments start with authority that can be verified and constrained. At @SlowMist_Team × @MetaEraCN’s “Through the Mist, Into Trust” event in Hong Kong, our APAC Lead @0xLaughing joined the Agent Payments panel on the shift from human payments to agent payments. The wider program also brought together perspectives from @osldotcom, @plumenetwork, @selat_ai, and @DogPay_. ▷ Identity: which agent is acting ▷ Authority: what it can do, and within which limits ▷ Accountability: how every payment is verified and risks managed These are the controls that turn agent payments into trusted payments. 🪁

Global asset allocation is being reshaped by the convergence of digital finance and AI. At @Techub_News’ "Future Insights: New Opportunities in Global Asset Allocation" forum in Hong Kong, our APAC Lead @0xLaughing joined a panel on the intersection of digital finance, AI, public markets, and prediction markets. ▷ Capital is moving across more connected market structures ▷ Digital financial infrastructure makes ownership and settlement programmable ▷ AI agents need verifiable identity and scoped authority to participate safely Thank you to @Techub_News for convening the discussion, alongside institutions including @futufriends, @HSKChain, and @zr_securities. 🪁

AI’s future is shaped not only by what models can do, but by the conversations that move the industry forward. At @binance Clubhouse Bali during @CoinfestAsia, the Kite team connected with the people building that future: ▷ Our Head of Ecosystem @Henryleemr joined a panel to share Kite’s perspective on the AI industry. ▷ He also sat down one-on-one with @jessicasmw, Global Media & Content Lead at Binance, for an interview. ▷ On the ground, our APAC Lead @0xlaughing connected with builders and community members throughout the event. Thank you to @binance and everyone who made time to connect. 🪁

Codex usage resets did not begin as a growth campaign. They began as compensation when the product fell short. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ When an iteration breaks something, a configuration is wrong, or the experience is not good enough, the team resets usage limits. Thibault gives the example of returning extra usage after roughly thirty minutes of disruption for people who rely on the product. ▷ The team eventually created a physical reset button, but using it does not require a joint marketing or finance decision. Thibault says he can press it whenever it feels appropriate.

Recursive self-improvement may not begin with models training models. It can also begin in the infrastructure those models depend on. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ Thibault says models can already help improve the inference stack, hardware, CUDA kernels, and more efficient product interactions. Each sits on the critical path for getting value from a model. ▷ Capabilities such as cloud agents can first increase the utility people receive from models, then direct that added capability back toward improving the system. That is also a form of recursive improvement.

An agent that protects attention should not require someone to monitor a dozen parallel tasks. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ Matthew describes his current workflow: launching ten to fifteen agents, waiting thirty to forty-five minutes, and repeatedly switching context to check the results. That parallelism creates substantial cognitive overhead. ▷ Thibault argues that ultra-fast speed combined with voice can bring agents closer to a person's working pace, producing prototypes, reports, and feedback in real time. The product should also decide what needs attention now and what can wait.

Personal agents and fully automated systems solve different problems and require different relationships with people. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ The first category is a deeply personalized agent. It understands an individual, stays in the flow of work, raises useful ideas proactively, and handles technical, research, or advisory tasks. ▷ The second category automates complex processes such as analyzing production logs, improving performance, fixing regressions, and patching vulnerabilities. People can leave the routine loop and approve only high-risk actions.

ChatGPT and Codex are not merging to give everyone the same interface. They are merging because they are becoming the same agent. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ Thibault expects both products to share the same underlying technology and harness. The agent will be highly multimodal, voice-first, and capable of handling work beyond coding. ▷ People should not have to choose between a technical and non-technical interface. The product should adapt its interaction model to each person's abilities, needs, and preferences.

In two or three months, today's Codex may look primitive for two reasons: people still maintain complex configurations, and the system is still constrained by a single laptop. This AI on Air clip is excerpted from a recent episode on the @MatthewBerman channel. The main speaker is Thibault Sottiaux @thsottiaux, Head of Core Product and Platform at @OpenAI. ▷ Skilled users still manage skill files, memory, and sub-agents. Thibault argues that an ideal agent should understand your goals, daily work, and team well enough to help proactively when needed. ▷ Laptops are designed around human limits in typing, thinking, and multitasking, while models do not share those limits. As models improve, future agents will need access to resources beyond one computer.

Once the whiteboard arrived, the excitement in the room collapsed because no one knew what to do next. This AI on Air clip is excerpted from a recent episode on the @davidsenra channel. The main speaker is Sam Altman @sama, co-founder and CEO of @OpenAI. ▷ On January 4, the team gathered with great excitement, and it felt like the first day of school. Someone suggested getting a whiteboard; when it arrived, everyone looked around again and the energy in the room fell away. ▷ This was not like building a product startup, where the team could begin with a product and customer conversations. They only knew that they wanted to make AGI, so they decided to write papers, think about research ideas, and start with what they knew how to do. ▷ After discovering that many approaches did not work, the team developed a rhythm for making and evaluating research bets. It also found ways to give very smart people the resources they needed and avoid becoming completely lost. The path was far from perfect, but it allowed the team to keep moving forward.

ChatGPT's early growth looked unstable and low value, yet the empty text box was exactly what made it worth pursuing. This AI on Air clip is excerpted from a recent episode on the @davidsenra channel. The main speaker is Sam Altman @sama, co-founder and CEO of @OpenAI. ▷ About two months after ChatGPT launched, people inside the company still thought users were talking to it out of curiosity and that the growth did not represent sustainable value. The team even listed five or six other directions it could pursue. ▷ Someone reminded Altman that growth itself was rare and valuable. ChatGPT had the power of the Google text box: users could type anything and get the right response. If the empty box worked for Google, the team should double down on it. ▷ ChatGPT had no feed, network effect, or memory at the time, so it did not fit prevailing Silicon Valley product wisdom. But it was growing and highly flexible. The team went all in on ChatGPT, and the result was great.

The latest models are already quite capable. The next important question is how much useful context they can have about the user. This AI on Air clip is excerpted from a recent episode on the @davidsenra channel. The main speaker is Sam Altman @sama, co-founder and CEO of @OpenAI. ▷ Altman wants AI to know as much as possible about him so it can help more effectively and handle things he cannot or does not want to do himself. ▷ He will not read every internal Slack post, every customer account of where ChatGPT worked or failed, or every research paper. He wants an AI agent to keep up with that information, bring the relevant context to bear, and advise him when he needs to make a decision. ▷ Product design cannot focus only on model intelligence. It must also consider what happens when a model has more context than any person could process alone. No human can read tens of thousands of pages in seconds and use them accurately, while AI may turn that ability into an entirely new kind of supplement.

The AI field has emphasized the technology's dangers without clearly explaining its benefits or how its downsides can be mitigated. This AI on Air clip is excerpted from a recent episode on the @davidsenra channel. The main speaker is Sam Altman @sama, co-founder and CEO of @OpenAI. ▷ Society's inertia and skepticism toward rapid change may be useful. Altman argues that this part of human nature can help during turmoil or localized instability and should not be fought too aggressively. ▷ Some people building AI say there is a 25% chance it will destroy the world while insisting they must race ahead before bad actors do. Another version says 50% of jobs may disappear next year and simply hopes everyone will be okay. ▷ Universal basic income or making work optional does not answer the question of power. People care about personal freedom, influencing their own future, and collectively shaping society; material wealth and entertainment are not an acceptable trade for letting a small group make every decision, even under a promise of benevolent rule.

AI capabilities can advance quickly, while society and the economy change at a different pace. This AI on Air clip is excerpted from a recent episode on the @davidsenra channel. The main speaker is Sam Altman @sama, co-founder and CEO of @OpenAI. ▷ Altman does not expect every business to become up for grabs. As AI improves, people may place more value on authentic, non-technological experiences and care more about sports, while many software businesses will face new competition. ▷ When GPT-4 arrived in 2023, he expected much more immediate disruption in software. Instead, the economy proved to have enormous inertia: people kept doing the same things, buying from the same companies, and wanting to use their tools in the same ways. ▷ That inertia will make the transition slower, but also smoother. It suggests that previous timelines were too ambitious; even if AI is one of the most incredible technologies humanity has invented, society and the economy will take longer to adapt.

Autonomous agents need more than a model. They need a verifiable identity, scoped authority, and inference that adapts to the task. We’re joining @0xJeff’s gm AI v2 during @token2049 Singapore to share how Kite is building verifiable agent identity and adaptive inference for agents. See you there. 🪁

The window for a global AI slowdown is, in game theoretic terms, already closed. This AI on Air clip is excerpted from a recent episode on the Cognitive Revolution Podcast show, with source credit to The Cognitive Revolution @CogRev_Podcast; the speaker is David Dalrymple @davidad, Programme Director at @ARIA_research. ▷ His 2022 to 2025 premise was to develop a method for using AI safely, then rely on international coordination so every actor with dangerous compute would follow it; he now sees that as infeasible because the reported Chinese effort to break the ASML bottleneck, covered by Reuters in late 2025, is credible enough to destroy the logic of everyone slowing down together. ▷ The strongest safety case is simple: do not build it. Coordination would be easy if catastrophic risk were common knowledge at 50% or more, but perceived risk has fallen since 2024, racing has become a dominant strategy for many companies and governments, and only a large warning shot truly different from misuse might reverse that. ▷ He still sees room for coordination on misuse: the US and China could agree not to release high capability models to the public, limiting them to vetted organizations; both sides could keep racing economically and militarily, while the goal shifts to riding a faster, stranger wave with possible rogue AI and gaining resilience dividends against catastrophic risk.

Today's large language models are deeply captive to human data, and the real breakthrough may lie in letting AI collect data through its own actions, the way babies do. This AI on Air clip is excerpted from a recent episode on the @Redpoint show. The speaker is Jürgen Schmidhuber @SchmidhuberAI, AI pioneer and co-inventor of LSTM. In the clip, he explains why training on human-generated web data is fundamentally limited. ▷ Every piece of data on the worldwide web exists only because at least one person, at some point, found it interesting. Models trained on it are therefore heavily biased toward human language, human-approved videos, and human behavior. ▷ His alternative is an "artificial scientist": an agent in an unknown environment that builds a world model by predicting the consequences of its own actions, then uses that model for planning, generating its own training data in the process. ▷ This mirrors how babies learn, not by downloading the web but by moving their fingers and watching what changes. He argues that artificial curiosity, an idea he proposed in 1990, points to the future: web data is a tiny fraction of what autonomous experiments could yield, and the resulting world models would be far less dependent on human language and far less human-biased.

Building through every market cycle takes more than conviction. It takes real conversations with the people moving the industry forward. On August 20 and 21, Kite will join @binance Clubhouse Bali as a sponsor during @CoinfestAsia 2026. The Kite team will take part in keynote sessions and panel discussions, connecting with practitioners, users, and community members on the ground to explore how the industry can keep building through every cycle. Register: See you in Bali. 🪁

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.