The Prompt

Hello reader,

Demis Hassabis, who has spent much of this year saying AGI may be close, stepped down as DeepMind CEO last week.

Wait, Demis. Aren't you supposed to be here to build AGI?

Not quite. The Nobel Prize winner isn't leaving the AI race - he will still serve as chairman of DeepMind and chief scientist of Alphabet. He also continues to lead Isomorphic Labs, Alphabet's subsidiary focused on accelerating drug discovery using AI.

Put simply, he's back to tinkering with AI and science, leaving the managerial stuff - we all know what that can do to creativity - to someone else.

Other reshuffles of humans within Google point to a similar trend. Jeff Dean, Google's 30th employee and former chief scientist, has left the company to establish Discovery Loop, a science-focused AI startup, with three other senior Google fellows.

With some of AI's biggest brains turning their attention to science, you could argue that chatbots have passed their novelty phase. Using AI for scientific discovery isn't new, but whether it can become a viable business remains unproven.

Hypotheses alone won't cut it. The test is whether AI systems can reduce the number of experiments, shorten development cycles, and produce discoveries that make it beyond the lab - the ChatGPT moment, if you will, but for science.

If science represents AI's long-term ambition, wearables represent its most immediate commercial opportunity. AI glasses, AI headphones, AI watches - you name it - these things are here already, but as my colleague Melissa reported in this piece, they still need a good reason to exist.

At least one player believes AI glasses will replace smartphones in the future. Others, like Razer, believe headphones are the more ideal form factor. We'll see how that goes over the next couple of years, as even Apple is reportedly working on both AI glasses and earphones.

I also wrote an opinion piece examining ByteDance's reverse cloud strategy, anchoring on one hit AI model. Its video generation model Seedance has attracted enterprise demand, but converting that into a wider customer relationship is a different story.

Whether in laboratories, on our faces, or inside the enterprise tech stack, AI is running into the same test: turning an initial breakthrough or burst of demand into something people keep using - and paying for.

Glenn Kaonang, tech journalist

TOP AI READS FROM OUR DESK

1️⃣ AI wearables chase payments, pets, and the perfect use case

Shipments of AI glasses more than quadrupled to 8.7 million units in 2025, with Meta controlling over 85% of the market. What do their customers use them for? In Japan, translation is gaining traction, while users in China can already pay for coffee by tapping the glasses' frames.

The category still has to deal with limited battery life, inconsistent accuracy, privacy concerns, and thin app ecosystems. For now, the hardware is arriving faster than a compelling everyday use case.

2️⃣ ByteDance has a billion-yuan wedge, but is it a cloud strategy?

Seedance reportedly generates over 1 billion yuan (US$147 million) in monthly revenue. Though the figure is unconfirmed, the commercial success gives ByteDance an opportunity to use the model to pull customers deeper into its broader enterprise offerings under BytePlus.

That's not easy in practice. Vietnamese AI startup Diaflow is a major Seedance customer, but it still keeps its infrastructure, databases, and core applications on Amazon Web Services rather than moving them to BytePlus.

WHAT WE THINK

1️⃣ ByteDance has too much too lose to play the distillation game

ByteDance founder Zhang Yiming reportedly told the company's AI team not to distill rival models to improve its own model capabilities, even if that means falling behind in benchmark leaderboards.

The company is playing it safe. Unlike most of its Chinese competitors like DeepSeek, Moonshot, and Z.ai, ByteDance has a consumer business in the US under intense scrutiny. Any accusation that ByteDance copied an American model could give the US more ammunition against TikTok.

ByteDance has always been an outlier in China's AI ecosystem. While other labs gain attention through open-weight models, the company has kept most of its flagship models proprietary.

It can afford that approach, though. ByteDance's chatbot app Doubao led China's consumer AI market with 382 million monthly active users in May, more than Qwen and DeepSeek combined. Seedance has also become an international breakout despite ByteDance trailing in the broader LLM race.

Better domestic chips may eventually reduce the need for Chinese labs to distill foreign models, though ByteDance may look the other way regardless. For most Chinese labs, falling behind is the bigger threat; for ByteDance, another TikTok-sized political crisis would be much worse.

2️⃣ A 10-million context window could shrink the AI agent stack

US-based AI agent startup Pokee, which we profiled last year, launched a model called Isaac. It comes with one standout claim: a context window of 10 million tokens.

A context window refers to the amount of information an AI model can consider at one time. For agents, that space fills up with instructions, documents, conversation history, and tool outputs. Once it reaches the limit, information has to be removed or compressed, which can result in poor execution.

AI agent startups work around this by building retrieval systems, summarizing earlier actions, discarding old tool outputs, and saving important details for the agent to use later. A larger context window lets an agent work across entire codebases and long interaction histories with fewer workarounds.

In practice, Pokee's latest model allows startups to spend less time squeezing information into the prompt and more time designing workflows and evaluations.

The catch is that Pokee demonstrated its context window and agent performance separately. Its 10-million-token result comes from a controlled internal test, while its agent benchmarks do not use the full context window.

Independent testing needs to show whether the larger context window actually makes agents better, rather than simply allowing them to take in more information.

STANDOUT AI DEALS

1️⃣ DeepSeek resumes fundraising at $74b valuation

DeepSeek is seeking close to US$8 billion at a US$74.1 billion valuation, with part of the money intended for a large data center in Inner Mongolia. The round is a reminder that efficient models such as DeepSeek's haven't made the frontier AI race cheap.

2️⃣ Moonshot AI targets $50b valuation in new round: sources

Moonshot's new target comes after a fundraise last month, which it closed at a valuation of around US$30 billion. It's also restructuring for a possible Hong Kong listing, which would further test public investors' appetite for China's costly AI model race.

3️⃣ A16z co-leads $300m round for SG AI firm Volta Infra

Volta Infra raised US$300 million at a US$2.4 billion valuation and secured another US$5 billion to finance customers' access to chips and computing capacity. The larger package shows that the AI infrastructure race increasingly runs on power, land, and complex financing - with overbuilding and aging hardware as the risks.

RESEARCH SPOTLIGHT

Can AI do AI research, or only the engineering work?

Researchers at Stanford, Princeton, UC Berkeley, and other universities gave AI agents six days and thousands of dollars in computing resources to do their own research. They were given the central question of two unpublished human-written research papers, and the agents were tasked to produce their own papers based on their findings.

The first paper asked whether an LLM's personality could be changed by adjusting the model itself rather than its prompts, and whether that would also change risky behavior. The second paper asked whether an AI system could recognize when unfamiliar data was making its predictions less accurate, before anyone knew the correct answers.

After writing the code and running experiments, the agents wrote two papers. However, the researchers behind the original studies said neither AI-produced paper was good enough for a top AI conference.

The study suggests that today's agents can carry out research tasks but still struggle to decide which ideas are worth pursuing or when an approach has failed. The next test is whether this holds across other research questions, and whether agents perform better when humans remain in charge of the research direction.

AI's push into science will depend as much on research judgment as on running experiments faster.

Catch up on past issues of the newsletter here. In the meantime, if you have any feedback or ideas, feel free to get in touch with Huong, our managing editor, at huong@techinasia.com.