AI Digest, Oct 6: Instinct's $10B Valuation, Agent Security, Huawei Supernodes

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers a $10 billion agent valuation with no disclosed revenue, how to secure autonomous agents, and a split in how supercomputers are built.

0:15Usually tech valuations have some sort of financial gravity, right?

0:20You build a product, you get users, you generate revenue, and the market calculates a multiplier.

0:26But if you look at the frontier AI landscape right now, that calculator is just completely broken.

0:33Today, we are exploring a world where AI agents are securing $10 billion valuations with literally zero revenue.

0:42Models are independently discovering physical materials.

0:45and there's this massive geopolitical split in how we actually build supercomputers.

0:50Yeah, it is wild.

0:51I mean, the market isn't pricing a product right now.

0:54They're pricing an entirely new paradigm of computing.

0:57The traditional SaaS metrics of, you know, monthly recurring revenue

1:01just do not apply when you're funding digital entities that autonomously take action.

1:06Right. And let's ground this with a specific one.

1:09Instinct.

1:09So it's a personal AI agent startup

1:11and they just raised a billion dollars at a 10 billion dollar valuation and this is only 33 days

1:18after their series B. Wow just 33 days. Yeah zero disclosed revenue and roughly like a hundred

1:25thousand users that is a hundred thousand dollars per user. It feels like buying the stadium before

1:31the sport is even invented. Like are investors just hallucinating here? Well they're placing a

1:37are all making a massive bet on action-driven agents, basically becoming the baseline infrastructure of the entire economy.

1:43Like, deep mine researchers are literally already forecasting a science economy.

1:48Wait, a science economy? What does that actually mean?

1:52So imagine a digital market where automated AI scientists are negotiating with each other to share scarce physical lab resources These agents will need a mechanism to trade compute time or lab access to run biological experiments

2:08While we sleep, basically.

2:10But wait, if these agents are out there trading resources and acting on our behalf, handing them the keys to our corporate networks feels incredibly reckless.

2:18A standard security clearance isn't going to cut it.

2:20No, it won't.

2:21And to solve that, Google Research just published this framework on contextual integrity.

2:27So we're moving away from like static security checkboxes.

2:31Okay, so what do they use instead?

2:33Agents are getting dynamic policy engines.

2:35The AI basically has to understand human social and professional norms.

2:40I mean, how does a machine understand a social norm?

2:42Give me an example of how that works.

2:44Right, so it evaluates the context of the request, not just the file permissions.

2:48The AI learns it's perfectly fine to email your daily meeting schedule to a co-worker, but if that same co-worker asks the AI for, say, your medical data or your salary history, the agent knows to block it.

3:02Oh, wow. Okay, that makes total sense.

3:04PM Lead ProFor a team lead: Google Research's contextual integrity means an agent judges the context of a request, not just file permissions. Salary or medical data requests from a co-worker should be blocked.

3:17And tracking what they actually generate is the other side of that coin, right?

3:21OpenAI is rolling out text watermarking to comply with the EU AI Act, but the industry pushback was always that watermarking inherently ruins the output quality.

3:31Yeah, but they actually proved it doesn't. On their Astra tests, the watermarked text scored a 49.76, which is compared to a 49.57 un-watermarked, so there's barely a difference.

3:43DeepMind is even taking it a step further with SynthID Bio, which is watermarking AI-designed biology.

3:49That is insane Yeah We building a verifiable digital paper trail for everything these agents create But all of this context awareness and autonomous action requires an insane amount of compute

4:02It really does.

4:03And looking at recent Chinese briefs, there is a fundamental split in how that hardware is being built.

4:09Huawei's Ascend chips have reportedly surpassed NVIDIA's market share in China.

4:13But they aren't just copying NVIDIA's homework, are they?

4:16No, far from it.

4:17because China faces strict limits on individual chip manufacturing.

4:21They're bypassing that bottleneck by innovating on the cluster architecture itself.

4:26So what are they doing differently?

4:27Huawei is pushing this Perium architecture and something called Unified Bus.

4:33They are linking up to 256,000 chips into one giant supernode.

4:40Wait, 256,000 chips acting as one brain?

4:44Yeah, it's a brute force scale strategy.

4:46And here's the real divergence. The West is pushing for co-packaged optics, or CPO, which puts the optical connections directly onto the silicon.

4:54Right.

4:55But Huawei is betting heavily on NPO, which is near-packaged optics.

4:59Okay, I need a translation on that. Why does moving the optics slightly off the chip matter?

5:04Well, because if a laser breaks on a co-packaged chip, you literally have to throw away the entire expensive processor.

5:10Oh, yikes.

5:11Right. With near-packaged optics, the connections are separate. So it offers vastly better cost and easier maintenance for these massive clusters. China is deliberately decoupling from Western hardware standards to build resilience.

5:24Well let see what is actually running on this global silicon with our model radar First up Vales AI used CLAUD 3 Opus to discover two new room anti semiconductors Which sounds like a total sci mouthful It really does But it basically means AI is successfully executing complex material science

5:43that usually takes humans years of trial and error.

5:46And if we look at inference efficiency, Reflection just released Beam 501B.

5:51It's a massive open-weight mixture of experts model. So it has 501 billion total parameters,

5:57but only 23 billion are active during any given task.

6:00Hold on, explain that mixture of experts part. How does ignoring most of the model make it better?

6:06Think of it like a massive university. You don't ask every professor on campus to solve a math problem, right?

6:11Right, you'd only consult the math department.

6:13Exactly. The AI routes the prompt only to the relevant expert neural networks within it,

6:18so it provides frontier-level logic for agentic coding while saving massive amounts of compute.

6:24That makes a lot of sense.

6:25Finally, we have Falcon Emirati 7B from Hugging Face and TII.

6:29It's a highly specialized 7 billion parameter model that masters the UAE dialect and cultural context.

6:36Which proves that local nuance is going to drive enterprise adoption just as much as for a horsepower.

6:41Definitely.

6:43PM Lead ProFor a company owner: model size isn't the only criterion. Beam activates 23 billion of its 501 billion parameters per task, and the hosts argued local nuance, like Falcon Emirati's, drives adoption too.

6:57Which brings us to a practical takeaway for any engineering lead listening to this.

7:02When your team is adopting AI, you need to evaluate its contextual integrity.

7:07Don't just test if your new agent can complete a complex coding task.

7:10Yeah, you need to aggressively test if it knows what not to share based on your company's internal norms.

7:17Test for the boundaries, not just the capabilities.

7:20I'll leave you with this to mull over.

7:22If AI agents start trading compute and lab time in their own machine economies, will human engineers eventually end up working for the agents?

7:31Something to think about as you watch this new sport being invented.

7:36PM Lead ProFind the transcript and sources on the episode page at pmlead.pro. Question for you: what would you test first to check your agent knows what not to share?