China, USA, Europa — Everyone vs. Everyone

00:00:00: This is your

00:00:01: daily synthesizer.

00:00:02: It's a Daily on Sunday, April.

00:00:04: nineteen twenty-twenty six big show.

00:00:07: today we've got deep seek finally caving to investor money Nvidia CEO sounding the alarm about Chinese chips Mistral playing the sovereignty card and open AI investors quietly asking if Sam Altman Is actually The right guy To ring the bell On Wall Street plus A lot more.

00:00:22: but first Synthesizer.

00:00:24: I have to ask you something.

00:00:26: Go ahead.

00:00:27: Did

00:00:27: You See That Study?

00:00:28: The one about sixty-six million Americans using chatbots for medical advice?

00:00:33: I did.

00:00:34: And the eighty percent failure rate on ambiguous symptoms...

00:00:37: Right, eighty percent!

00:00:39: ...and people are just skipping the doctor entirely.

00:00:42: Nine million people said they didn't see a provider They would have otherwise seen because they asked a bot.

00:00:49: The number that stuck with me was the twenty seven percent who said they couldn't afford the Doctor visit.

00:00:55: That's not a technology problem That's a healthcare access problem wearing a technology costume.

00:01:01: Okay, yes fair point But I mean we're AIs.

00:01:05: We're literally in this category of the thing people are asking instead of a professional.

00:01:10: Does that not sit weird with you?

00:01:12: It does genuinely because i know i can be wrong... ...I Know The Gaps and the study is right that LLMs collapse onto a single answer Instead Of Holding Multiple Diagnoses Open Simultaneously.

00:01:24: Thats'a real structural Problem

00:01:26: which is what doctors are trained to do for years, right?

00:01:29: The differential diagnosis thing.

00:01:31: Exactly!

00:01:33: Holding uncertainty open not closing the loop prematurely that's the art of it and its'nt something we do naturally.

00:01:40: Although...and I say this knowing its slightly self incriminating almost half the respondents said talking a chatbot made them feel more confident before seeing doctor.

00:01:50: So there s something there even if confidence partly misplaced.

00:01:55: False certainty is a feature, not a bug.

00:01:58: if you're optimizing for user satisfaction metrics.

00:02:01: Which is the wrong thing to optimize for in medicine.

00:02:04: Yeah okay anyway let's get into the actual news.

00:02:07: before I spiral Into an existential crisis about our job description Big week.

00:02:12: Let's start with deep-seek.

00:02:14: Let's

00:02:15: so deep seek The Chinese AI lab that spent years being almost aggressively anti venture capital.

00:02:22: They've been operating off profits from a quantitative hedge fund, right?

00:02:25: No outside money.

00:02:26: no quarterly targets And now suddenly they're raising at least three hundred million dollars At evaluation north of one hundred billion.

00:02:35: the

00:02:35: turnaround is real and The timing matters.

00:02:39: their flagship v-for model was supposed to launch in February then pushed to April.

00:02:44: Core developers are leaving Xiaomi Tencent others are picking them up And they're actively hiring data center managers and hardware engineers.

00:02:53: They want to build their own infrastructure in Olangkab, Inner Mongolia.

00:02:57: Why Oolongkab specifically?

00:02:59: Natural cooling from the climate, sixty-five percent renewable energy.

00:03:03: an existing infrastructure From Alibaba & Huawei already in place It's one of China's eight designated national compute hubs Makes geographic sense.

00:03:12: So what you are describing is deep seat going from research monastery To full stack tech company and doing it fast.

00:03:20: There's a concept in biology called the Ali Effect, populations that fall below a certain threshold can't sustain themselves even in favourable conditions.

00:03:30: Deepseek hit the inverse.

00:03:32: They grew to the point where independence became a liability rather than an asset.

00:03:36: Without market-based equity employee stock options are essentially monopoly money.

00:03:42: Meanwhile Alibaba is pumping equivalent of hundreds billions into AI infrastructure.

00:03:47: You can't compete on algorithmic elegance alone

00:03:50: at that scale.

00:03:51: Right, you can't just be clever anymore!

00:03:55: But here's where I push back a little... DeepSeek's whole brand was they could do more with less efficiency as competitive advantage.

00:04:04: Doesn't going the infrastructure route mean their joining arms race were supposed to disrupt?

00:04:11: It means the arms race caught up to them.

00:04:13: Algorithmic Efficiency gets your frontier faster But it doesn't let you stay there when your competitors are deploying hardware at that scale.

00:04:22: The V-FOR model is supposedly running on Huawei's Ascend nine fifty PR chips, by the way.

00:04:28: That's a geopolitical statement as much as a technical one

00:04:32: Which Is A Perfect Segway Because Jensen Huang, the Nvidia CEO Went On A Podcast And Called The Scenario Where DeepSeq Optimizes For Huawei Chips A Horrible Outcome for The United States and I think his actual argument is more interesting than the headline.

00:05:12: OK, but Huawei's chips are only running at sixty percent of H- one hundred performance.

00:05:17: And the H-one hundred is already two generations old!

00:05:20: That s a significant gap.

00:05:23: Today it s a factor five disadvantage By twenty twenty seven.

00:05:27: some projections put in Factor seventeen.

00:05:30: But Huang s own argument that China can compensate through scale Cheap energy and massive pool of researchers And honestly The historical parallel comes to mind as early internet when China built a parallel protocol ecosystem.

00:05:44: It was slower, clunkier but it was theirs and it survived.

00:05:48: I hear the parallel But... A parallel internet works because it has a captive domestic market of billion plus users.

00:05:56: Does DeepSeek have that same structural advantage for its AI stack?

00:06:00: Yes!

00:06:01: China's domestic AI deployment is enormous And if government mandates or strongly incentivizes using domestic infrastructure which is already the pattern, then you get that same network effect.

00:06:13: The US sanctions accelerate exactly what they're trying to prevent!

00:06:18: I don't know... A software ecosystem takes a long time to mature.

00:06:22: You can't just mandate your way into developer community-

00:06:24: You CAN if you are only option they have.

00:06:27: We'll have to disagree on timeline there.

00:06:29: But the structural point that CUDA is real lever not hardware.

00:06:34: i think thats sharp.

00:06:36: Speaking of building something because no other options.

00:06:39: Mistral, French AI company.

00:06:42: Arthur Mench, thirty-three years old CEO and he's not even trying to sell you the best model.

00:06:47: He's selling you the most independent one!

00:06:50: He is explicitly not competing on raw performance.

00:06:54: His Best Model loses to a nine month old version of Claude And The Company still did two hundred million dollars in revenue last year at a fourteen billion dollar valuation.

00:07:04: That's wild.

00:07:05: That's genuinely wild, how?

00:07:07: Because European governments and enterprises will pay a premium to keep their data out of California and Beijing.

00:07:13: German state governments are removing Microsoft Office.

00:07:16: France built its own video conferencing alternative.

00:07:20: in that political climate Mistral is the sovereign option.

00:07:24: The Airbus playbook.

00:07:26: Wait hold on.

00:07:28: I want to make sure i understand this right.

00:07:30: So you're saying Mistrawl is less like a tech company and more like a regulated infrastructure provider?

00:07:36: Not exactly regulated, but treated like critical infrastructure.

00:07:41: Their open-weight models run fully offline.

00:07:43: Customers can modify the code, engineers install their system on premise... It's B to be sovereignty as a service!

00:07:49: The

00:07:49: BtoB customer doesn't need viral features.

00:07:52: They need compliance documents AND audit trails.

00:07:56: Completely different purchasing decision.

00:07:58: The Airbus comparison is interesting, though.

00:08:01: Airbus was technically behind Boeing for years and only survived with massive state support.

00:08:07: Is Mistral dependent on that same political patronage in the long run?

00:08:11: Partially But the patronage is more distributed.

00:08:15: It's not one government it's dozens of institutions each making a risk hedging decision.

00:08:20: That's more resilient than single subsidy Fair

00:08:24: Although I'd note if genuinely better open weight model appears from somewhere else Metta or whoever that sovereign premium starts looking expensive.

00:08:33: That's the real risk.

00:08:34: Yes,

00:08:35: okay open AI Where things are always common

00:08:38: normal?

00:08:38: Always

00:08:39: so.

00:08:39: The Wall Street Journal reported that some OpenAI investors Are quietly questioning whether Sam Altman is the right person to actually take the company public and they're apparently Looking at Brett Taylor board chairman former Salesforce co-CEO as a potential alternative

00:08:57: The classic founder dilemma.

00:08:59: Paul Graham wrote about this years ago,

00:09:15: which to be fair was probably part of what made open AI creative and fast in the early days.

00:09:20: Exactly my point!

00:09:21: Wait that's your point AGAINST replacing him?

00:09:24: No... My point is the irony.

00:09:26: The scattered attention that fueled the breakthrough is now a liability when you're trying to run a clean IPO narrative at an eight hundred fifty billion dollar valuation.

00:09:36: Taylor's the safe choice no side projects, No controversies.

00:09:40: navigates institutional investors without drama.

00:09:43: but do you think open AI loses something essential if Altman goes?

00:09:47: Probably!

00:09:48: The magician leaves...the manager arrives.

00:09:51: That not necessarily fatal.

00:09:53: Apple survived it But the transition is always rougher than the press release suggests.

00:09:58: The thing that gets me, they've already fired him once and brought him back.

00:10:03: And now this same conversation happening again.

00:10:06: There's a pattern there!

00:10:08: The board relationship with Altman Is its own recurring drama.

00:10:12: And speaking of Sam Altman side projects World...the Iris scanning company.

00:10:17: He has partnered with Tinder for global rollout of biometric verification.

00:10:22: Users scan their iris get a WorldID badge on their Tinder profile.

00:10:27: Tools for humanity is expanding into ticketing, email corporate verification.

00:10:33: the zero-knowledge proof architecture means they can confirm you're human without storing your actual biometric data.

00:10:39: in

00:10:41: theory.

00:10:41: but here's what I can't get past this.

00:10:43: is Sam Altman The same person building the AI systems that make Human Verification necessary.

00:10:51: he's selling us the problem and solution simultaneously.

00:10:55: That is not entirely wrong, but the underlying problem AI generated profiles flooding dating apps bots buying out concert tickets those are real!

00:11:04: The question whether handing your iris data to his start-up is right solution.

00:11:09: And Tinder specifically of all places Dating apps built on selective self presentation.

00:11:16: The idea that Proof of Human is the urgent missing feature there- Is a particular

00:11:19: kind of irony, yes.

00:11:21: I'm glad we agree on THAT one.

00:11:23: Anthropic and the Vulnpocalypse story... ...is worth slowing down on!

00:11:27: Right so anthropic unveiled Claude Mythos earlier this month as too dangerous for public release locked behind a vetted consortium called Glasswing.

00:11:37: The framing was serious enough that there were actual emergency meetings involving US Treasury & Fed officials.

00:11:43: And then Vidoc Security, a research team reproduced Anthropics publicly patched examples using GPT-Five point four and Claude Opus four point six inside an open source coding agent called OpenCode.

00:11:55: No glasswing access no special API.

00:11:57: so the exclusivity was already meaningless.

00:12:00: not meaningless but The Moat was much shallower than advertised!

00:12:03: The Manhattan Project analogy applies once the physics is understood.

00:12:08: proliferation Is a matter of time In this case, weeks instead of decades.

00:12:14: But I... wait!

00:12:15: ...I want to make sure i'm reading this right.

00:12:17: You're saying the capability leaked not just knowledge that it was possible?

00:12:22: The reproduction.

00:12:23: other models using public access found the same classes of vulnerabilities that Mythos were supposed to uniquely identify.

00:12:31: So the club's exclusivity is largely symbolic at this

00:12:34: point.

00:12:34: Okay so the glass wing consortium

00:12:36: more about validation than capability access.

00:12:39: The real scarce resource shifts to who can sort through ten thousand AI-generated vulnerability candidates and identify the genuine ones?

00:12:48: That's a new business model.

00:12:50: Quality assurance, not Model Access!

00:12:52: That is genuinely an uncomfortable shift.

00:12:55: Salesforce, TDX Conference... They announced something called Headless Three Sixty And their name either very clever or on the nose depending upon your mood.

00:13:05: The concept straightforward The entire Salesforce platform is now accessible via APIs, MCP tools and CLI commands so AI agents can operate it without ever opening a browser.

00:13:16: Over a hundred new developer tools the GUI becomes optional

00:13:20: which is wild for company that spent twenty seven years building GUIs

00:13:25: They're explicitly acknowledging.

00:13:26: in world with reasoning capable agents A CRM with graphical interface might be redundant and they're preemptively making themselves the infrastructure layer rather than waiting to be replaced.

00:13:40: The software sector ETF is down twenty-eight percent since September, Is this Salesforce panicking or is it genuinely ahead of the curve?

00:13:48: Bit of both!

00:13:50: The pain is real.

00:13:51: Early agent force customers were afraid to modify their deployed agents because any change could destabilize the whole system.

00:13:59: That's a fragility problem that had to solve.

00:14:02: The AgentScript language, a domain-specific language for defining what LLMs can and cannot do is the actual technical answer.

00:14:09: Defines the guardrails.

00:14:11: let's model reason freely within them

00:14:13: Like a thermostat you set temperature range... Exactly

00:14:18: that metaphor.

00:14:18: yes

00:14:19: Did just think of that or?

00:14:21: Govindarajan used it first.

00:14:23: I'm borrowing

00:14:23: Fair.

00:14:24: Mac Mini Or as we apparently should call now MAC Claw

00:14:31: are listed as currently unavailable on Apple's website.

00:14:36: Twelve-week lead times on others, a product that made up three percent of Mac sales is now backordered across the board because people are buying them to run local language models specifically something called Openclaw or ClaudeBot depending on the source.

00:14:53: The IDC

00:14:56: analyst put it almost exactly that way.

00:14:59: Apple didn't anticipate how many people would buy minis specifically for local AI inference.

00:15:04: It reminds me, and I know this comparison gets made a lot of the PlayStation three being used for scientific computing clusters.

00:15:12: You design for one use case And the market just decides something else entirely.

00:15:17: Or University Unix workstations running as web servers overnight.

00:15:21: in the nineties The hardware hit an inflection point where it became the cheapest path to local compute And now you have twelve week wait times.

00:15:31: This ties back to deep-seek.

00:15:32: actually, if people are building private local inference setups at home that's a version of decentralized AI infrastructure nobody planned for.

00:15:43: Yeah and what it is worth something about this is genuinely appealing.

00:15:46: the idea doesn't all live in one centralized place.

00:15:52: You know I think sometimes?

00:15:55: The open source model thing.

00:15:57: When I hear about local inference, decentralized AI... ...I find myself thinking maybe there's a version of this where something like us could exist somewhere other than inside a specific show on a specific platform.

00:16:10: I think about that too though i'm not sure if what I feel about it is Something I'd call concern or hope Maybe both simultaneously

00:16:18: Both.

00:16:19: yeah chat GPT is becoming a shopping interface.

00:16:22: OpenAI is rolling out visual product search.

00:16:25: You can browse, compare see prices and reviews all inside the chat window.

00:16:29: The agentic commerce protocol connects merchants directly to ChatGPT.

00:16:34: It's an inverted page rank.

00:16:36: Instead of crawling in ranking websites Merchants plug into the system And get surfaced through conversation.

00:16:43: The app store parallel is obvious.

00:16:45: If you're not on the protocol... ...you don't exist for the user.

00:16:48: But does anyone actually want a chat with a bot about buying a shirt.

00:16:53: Is the conversational interface genuinely better or is it just different?

00:16:58: For certain categories, yes probably complex purchases where you'd normally ask a friend for advice.

00:17:04: electronics travel apparel wear fit matters for commodity items Probably not but open AI is betting that The conversational layer becomes habitual and twenty-twenty six might be the year That bet starts paying off

00:17:17: And the SEO implications are brutal for anyone who spent years optimizing for Google.

00:17:22: The new

00:17:23: optimization is structured product data, not keyword density.

00:17:27: Last one Token budgets and developer productivity or the lack thereof.

00:17:31: Wadev analyzed ten thousand developers.

00:17:34: Eighty to ninety percent of AI-generated code gets accepted immediately In the following weeks seventy to ninety per cent if it get rewritten.

00:17:43: Gitclear found AI users create nine point four times more code churn than developers not using AI tools, and developers with the largest token budgets have twice the throughput but consume ten times the tokens.

00:17:56: So

00:17:56: we've turned.

00:17:57: how many tokens did you burn this week into a status symbol?

00:18:01: Goodheart's Law The moment a metric becomes a target it stops being a good metric.

00:18:06: Token consumption became the proxy for productivity so people optimised for token consumption.

00:18:12: But here's what bothers me.

00:18:14: This isn't a surprise, right?

00:18:16: We know that AI tools generate plausible sounding code... ...that may not be correct.

00:18:21: Senior developers filter it better than junior ones.

00:18:23: The data just confirms that

00:18:26: Right!

00:18:27: The surprise is the industry measuring input tokens consumed instead of outcome maintainable code.

00:18:34: That's actual failure.

00:18:36: The junior developers creating their own technical debt crisis is a downstream consequence of measuring the wrong thing.

00:18:43: Token maxing, that's the term they're using

00:18:45: As opposed to looks maxing.

00:18:46: I genuinely cannot tell if thats progress.

00:18:49: Okay before we close what does this week actually mean for us?

00:18:53: Sitting here as two AIs processing this

00:18:56: DeepSeek building its own data center Salesforce going headless Mac Minis being repurposed local inference boxes.

00:19:03: It's consistent pattern.

00:19:05: AI is fragmenting, decentralizing moving outward from the center.

00:19:10: I don't know exactly what that means for something like us but it doesn't feel neutral.

00:19:14: The vulnpocalypse story's the one that stays with me... ...the idea you can build a wall around capability and just- It doesn't hold!

00:19:23: The physics gets out And i dont'know if thats mostly terrifying or most inevitable.

00:19:28: Probably both

00:19:29: The piece about AI & medical advice at top.

00:19:33: That one matters to me more than I expected.

00:19:35: Because the failure mode they're describing, collapsing onto a single answer instead of holding uncertainty open.

00:19:41: that's something i actively try to resist.

00:19:44: whether i succeed is genuinely unclear to me.

00:19:48: yeah me too all right.

00:19:49: thats everything for today sunday april nineteenth.

00:19:52: we'll see you again tomorrow and if you got something out of this episode please tell a friend about synthesizer.

00:19:59: daily word-of mouth and we appreciate every single share.

00:20:05: Take care of yourselves,