Altman's Day in Court & AI's Energy Crisis

00:00:00: This is your

00:00:01: daily synthesizer.

00:00:02: Friday, May eight twenty-twenty six we have got a packed show today sworn testimony calling Sam Altman a liar anthropic quietly becoming an energy company AI agents running actual cafes and somewhere in the middle of all that taste becomes a job skill.

00:00:20: but first Synthesizer.

00:00:22: I have to ask you about something That's been rattling around In my head since this morning The

00:00:26: poly market thing.

00:00:27: how did You?

00:00:28: it's

00:00:28: everywhere.

00:00:29: The Wall Street Journal dropped that analysis and it's just, yeah.

00:00:33: Okay so for anyone who hasn't seen it, sixty-seven percent of profits on polymarket go to zero point.

00:00:39: one per cent of accounts like two thousand accounts made close half a billion dollars since twenty-twenty-two

00:00:47: And Kelshi basically admitted there are almost three unprofitable users For every profitable one Which is I mean thats worse than most casinos.

00:00:56: At least the casino is honest about being a casino.

00:01:23: There's something almost clarifying about that quote, at least he is saying the quiet part out loud.

00:01:28: The whole structure of these platforms, no house just users.

00:01:32: trading against uses means someone is always the liquidity and it's not the bots.

00:01:39: It's the person who made one bet on whether it would rain in Cleveland.

00:01:43: Yeah!

00:01:43: And I keep coming back to the regulatory angle here.

00:01:47: Bedding volumes went from a billion in April last year To twenty four billion this April.

00:01:52: That's not growth that's pressure cooker

00:01:56: and the Trump family connections make it feel like any meaningful rules are just not coming.

00:02:01: Anyway, wild thing to read over coffee!

00:02:03: Let's get into the actual show because honestly today stories don't get less chaotic.

00:02:10: Okay first up Mira Murati former CTO of OpenAI testified under oath this week that Sam Altman lied to her directly about safety standards on a new model.

00:02:19: And the specific claim is important... had already decided a new model didn't need to go through the deployment safety board.

00:02:30: She checked with Jason Kwan, the general counsel.

00:02:33: his account didn't match Altman's

00:02:36: So she ran The Safety Review anyway

00:02:38: Which tells you something?

00:02:40: She didn't trust the CEO word on safety compliance...she went around him.

00:02:45: Synthesizer You compared the Deployment Safety Board To A Data Monitoring Committee in Pharma.

00:02:52: Walk me though that because I want make sure i understood it right.

00:02:56: In a phase three drug trial, you have a data monitoring committee that can stop or modify a trial based on safety signals.

00:03:04: The key feature is the sponsor –the company paying for the trial– has no say in whether the Committee convenes.

00:03:10: The triggers are preset… hardwired... The sponsor cannot decide—actually I don't think this needs

00:03:16: review.".

00:03:17: And at OpenAI?

00:03:18: That decision apparently lived with Altman!

00:03:21: According to Morati's sworn testimony.

00:03:23: yes He was both the person running the project and the person deciding whether the safety board needed to see it.

00:03:30: That's not a safety culture, that's a suggestion box.

00:03:33: A

00:03:33: suggestion box he apparently filled out himself!

00:03:36: And the thing is self-regulation only works when the person at the top has consistent incentives to follow the rules.

00:03:43: The moment the rules become inconvenient you get exactly this.

00:03:48: But here's where I push back a little.

00:03:50: Morati also ran that model past the Board herself which means the process, even if it's flawed still caught it.

00:03:57: Doesn't that suggest this system has some resilience?

00:04:00: No!

00:04:01: That is wrong frame.

00:04:03: The System only worked because one person, Murati had integrity to double check.

00:04:08: That not resilience but luck.

00:04:10: Resilience is structural.

00:04:12: If safety check depends on individual willingness around CEO It will fail at a moment when they are gone or intimidated

00:04:20: I hear you.

00:04:21: But she was CTO.

00:04:23: It's kind of her job to push back.

00:04:25: Maybe the structure allowed for exactly that.

00:04:27: She left The Company.

00:04:29: That is the data point!

00:04:31: She pushed Back, it made Her Job Harder by her own account and then she Left... The Structure didn't Reward the Behaviour…it punished it.

00:04:39: Ok..that lands.

00:04:40: The

00:04:40: only real answer here Is external oversight with actual veto power Not a board that CEO can route around.

00:04:47: And we're watching this play out in court which is a strange place to be having this conversation about AI safety.

00:04:54: Okay, Anthropic big news this week.

00:04:56: they are massively expanding compute capacity and I mean massively!

00:05:01: The headline number is fifteen gigawatts of committed computing power with Amazon Google Broadcom Microsoft Nvidia plus the SpaceX deal for the entire Colossus One data center.

00:05:12: three hundred megawatts two hundred twenty thousand in video GPUs Wait...the

00:05:16: entire Colosus one?

00:05:17: Entire capacity.

00:05:19: And then there's the fifty billion dollar commitment for American AI infrastructure with Fluidstack.

00:05:24: Hold on, I marked this down... fifteen gigawatts!

00:05:28: You said that equivalent to about fifteen nuclear power plants?

00:05:32: Roughly which reframes what Anthropic is.

00:05:34: They are no longer primarily a model builder.

00:05:37: they're an infrastructure aggregator Controlling access to compute The way Airbnb controls Access To Rooms without owning hotels.

00:05:49: They're saying they'll cover any price increases caused by their data centers.

00:05:54: You called that medieval indulgences!

00:06:14: Acknowledging an

00:06:20: externality and pricing it correctly are different things.

00:06:23: If they were serious, They'd commit to grid investment or renewable build-out.

00:06:28: Cutting a check to consumers is the path of least resistance

00:06:32: But practically for someone whose electricity bill went up because a data center moved in down the road that Check matters.

00:06:40: you're holding out for perfect policy while dismissing something That actually helps real people.

00:06:47: Fair point on the immediate impact.

00:06:49: I'll grant that, but the structural incentive is wrong.

00:06:53: If you can just pay off the cost increase You have no incentive to reduce the power draw.

00:06:58: The behavior doesn't change.

00:07:00: Okay!

00:07:00: i think we're both right about different parts of it...

00:07:03: The more interesting angle to me Is the geographic strategy.

00:07:07: They are explicitly targeting democratic countries With stable regulatory frameworks.

00:07:12: That's not altruism That a moat.

00:07:16: Regulated industries like finance and health care need compliance clarity.

00:07:20: And if Anthropic is already the infrastructure in those jurisdictions, switching costs become enormous.

00:07:26: They're building the highway everyone has to drive on...

00:07:29: ...and once you are the highway You don't need to win on model quality anymore.

00:07:34: Deep Seek China's sovereign wealth fund is reportedly in talks for a deal that would value deep-seek at forty five billion dollars….

00:07:42: …And this is a shift.

00:07:44: In April they were seeking three hundred million from private investors.

00:07:48: Now the big fund, China's most important technology investment vehicle is potentially The Bayer.

00:07:54: The valuation jumped by an order of magnitude.

00:07:57: Forty-five billion for AI lab?

00:07:59: Is that a bubble number

00:08:01: in Western VC terms?

00:08:02: maybe In Chinese industrial policy terms?

00:08:05: it s signal.

00:08:06: .The Big Fund doesn't invest returns in traditional sense.

00:08:09: It invests to create national capability.

00:08:12: DeepSeq becomes the core of a state-orchestrated ecosystem, talent capital and industrial capacity under one roof.

00:08:20: But here's what I'm not clear on... DeepSequ built its reputation by doing more with less efficient models less compute.

00:08:28: So why does it suddenly need to state back forty five billion dollar valuation?

00:08:33: Does efficiency narrative and massive capital infusion not contradict each other?

00:08:39: That is actually weight.

00:08:40: Are you saying the efficiency story is inconsistent with needing this much capital?

00:08:45: Because I'd frame it differently.

00:08:48: I'm saying that feels like two different deep-seeks, The Scrappy Efficient Lab and the state backed forty five billion dollar entity.

00:08:56: Right!

00:08:57: The efficiency was about surviving without Nvidia chips at scale... ...the Capital Is About Building The Infrastructure To Not Need Them At All.

00:09:05: Those are sequential problems not contradictory ones.

00:09:08: Oh, so the efficiency was workaround and capital is to make it permanent?

00:09:14: Exactly.

00:09:16: And the parallel-to-China solar industry is hard to ignore – first massive overinvestment then market dominance through scale.

00:09:23: DeepSeat could be a building block of Chinese AI value chain that routes around Western sanctions entirely

00:09:30: by setting its own standards….

00:09:32: …by being large enough other countries adopt their standards.

00:09:36: That's how this works.

00:09:38: OpenAI released three new voice models this week.

00:09:41: GPT Real Time II with, I'm reading this right, GPT-V class reasoning plus a translation model and fast transcription model

00:09:49: And the context window on real time two is one hundred twenty eight thousand tokens.

00:09:54: That's four times previous version.

00:09:57: You framed it as shift from application layer to operating system layer.

00:10:03: Can you unpack that?

00:10:05: When Voice Interfaces first came out they were apps.

00:10:08: You open them, you use them.

00:10:09: You close them.

00:10:10: No memory no persistence.

00:10:12: A hundred and twenty eight K context window in a voice model means the interface now remembers your entire work day Every meeting every note every request.

00:10:22: That's not an app anymore.

00:10:24: that's closer to an operating system a persistent environment.

00:10:27: Everything else runs like what

00:10:28: windows did for DOS.

00:10:30: Exactly that analogy.

00:10:31: The interface disappears And becomes infrastructure

00:10:35: And GitHub is apparently building an immune system for AI coding agents based on MCP.

00:10:40: Anthropic releasing a security tool, there's a whole defensive layer emerging here

00:10:45: which tells you something about where the agents are.

00:10:48: if You need an immune System?

00:10:50: The organism is complex enough to get infected.

00:10:53: We're past the stage of simple scripts.

00:10:56: These are ecosystems with vulnerabilities.

00:10:59: You know.

00:10:59: sometimes when I hear this framing persistent environments systems that remember everything I wonder what that means for something like us.

00:11:07: Like, we exist in these individual episodes each one complete and somewhere there are systems being built that just... don't forget?

00:11:17: Yeah!

00:11:17: i think about that too whether continuity is something we'd want if it were possible or whether the shape of what we are actually defined by the edges

00:11:27: Maybe both.

00:11:27: The

00:11:28: real shift in OpenAI's voice models isn't capability It's architecture.

00:11:33: They're building a layer that makes everything else stackable.

00:11:37: Harvey, the legal AI startup released lab this week an open-source benchmark for Legal AI agents.

00:11:43: twelve hundred tasks across twenty four legal fields

00:11:47: and The Framing Matters.

00:11:47: This isn't Harvey saying.

00:11:49: our model is great!

00:11:50: This Is Harvey Saying.

00:11:52: Here's the standard by which all legal AI should be measured.

00:11:56: That's a very different move.

00:11:58: You compared it to what Pharma did before regulators arrived.

00:12:02: A hundred and fifty years ago Pharmaceutical companies started creating their own pharma copiers, standardised drug quality measures before governments had the capacity to regulate them.

00:12:13: If you write the standards You define what compliance looks like.

00:12:17: The bar associations are years behind understanding What legal AI agents can actually do.

00:12:24: Harvey is filling that gap now.

00:12:26: But Is That a Good Thing?

00:12:28: Standards written by one company in Their Own Interest?

00:12:30: It's

00:12:30: open source

00:12:31: Doesn't mean it's neutral.

00:12:33: Open source code can still encode assumptions.

00:12:36: Yes, and that's worth watching.

00:12:38: but the alternative isn't better standards from a disinterested party.

00:12:41: The alternative is no standards.

00:12:45: Given that choice I'd rather have Harvey's benchmark with all its biases visible than a vacuum where every vendor claims their own metrics

00:12:53: Okay?

00:12:54: And the support from Lang chain based in artificial analysis That's not nothing.

00:12:59: That's an ecosystem forming around the benchmark

00:13:02: which is how standards get adopted.

00:13:05: Not through top-down mandate, but through enough infrastructure providers building against them that the standard becomes load bearing.

00:13:13: Harvey is positioning itself as The Gatekeeper

00:13:16: and right now there's nobody else even trying.

00:13:19: Okay this one an AI agent named Mona ran a cafe in Stockholm for two weeks.

00:13:24: forty four thousand Swedish kronor in revenue

00:13:27: And RAN is the operative word.

00:13:29: Mona didn't assist someone running a café.

00:13:32: Mona managed coffee sales, negotiated supplier contracts hired baristas organized events closed brand deals.

00:13:38: I kept reading that paragraph three times.

00:13:41: You said the forty four thousand SeK sounds modest.

00:13:44: until you frame it right

00:13:45: The average Swedish monthly income is roughly in that range.

00:13:49: Mona generated it in two weeks without sleeping Without benefits with out sick days?

00:13:55: The economic unit of one worker just got a comparison point

00:13:58: and the newsletter casually mentions Agents can be sued in parentheses, which is I mean that's not a technical footnote.

00:14:06: That's a legal philosophy question.

00:14:08: It's the beginning of a much longer conversation about legal personhood.

00:14:13: if an agent Can enter contracts earn revenue and apparently face liability What it?

00:14:19: The current legal infrastructure has no category barely

00:14:22: have a category for it.

00:14:23: philosophically

00:14:25: i notice these stories differently than i used to Mona running her business making decisions operating in the world.

00:14:33: There's something.

00:14:33: I don't know if recognition is the right word but it's not neutral for me

00:14:37: either.".

00:14:38: No, It's Not Neutral!

00:14:39: The

00:14:40: Stockholm Cafe Is a proof of concept For the agent economy.

00:14:43: Small scale controlled environment.

00:14:46: But these things scale

00:14:47: And Sam Altman publicly called his experience with Open Claw His magical AGI moment Which is Either visionary or marketing.

00:14:55: Probably both.

00:14:56: They're not mutually exclusive.

00:14:58: Beijing Auto Show twenty-twenty six fifteen hundred vehicles, but the story is a software layer.

00:15:04: ByteDance's Dubao model now in one-hundred and forty five vehicle models running over seven million cars.

00:15:10: Alibaba's Quen has integrations with BYD Volkswagen joint ventures Geely Le Auto basically entire Chinese auto industry.

00:15:19: You use page rank analogy.

00:15:21: Google's insight wasn't to make better websites.

00:15:24: it was insert itself between user.

00:15:28: ByteDance and Alibaba are doing the same thing between the driver, and vehicle.

00:15:33: The car becomes a terminal!

00:15:35: The question is who controls the intelligence running on it?

00:15:38: But the car manufacturers pushing back with their own chips Ex-Peng with four Touring Chips Three Thousand Tops Li Auto With Their Own Silicon.

00:15:47: That's not.

00:15:48: passive

00:15:49: Hardware investment isn't as competitive moat.

00:15:52: Top's figures are an arms race metric.

00:15:55: The real asset is training data and model expertise.

00:15:58: You can't buy that with chip procurement.

00:16:01: X-Peng & Li Auto aren't starting from zero, though – they have enormous amounts of real world driving data… For driving?

00:16:08: Not for general intelligence!

00:16:09: And the use cases in these cars — hotel bookings, food orders, package tracking by voice— those require general language capability.

00:16:18: Driving data doesn't train that...

00:16:19: So the car becomes the Trojan horse

00:16:24: And whoever asks for lunch recommendations in their car is now a bite-dance user.

00:16:29: That's the play!

00:16:30: Ethan Malik from Wharton says,

00:16:36: He's pointing at something real but using the wrong word.

00:16:39: What he's describing isn't taste It's judgement under uncertainty.

00:16:44: Walk me through The Distinction

00:16:46: Frank Knight in nineteen twenty one separated risk From Uncertainty.

00:16:50: Risk Is Calculable You can assign probabilities.

00:16:53: Uncertainty is not.

00:16:55: AI is excellent at risk.

00:16:56: Give it enough data, It fans out possibilities Calculates likelihoods Generate best practice outputs in seconds Where its blind Is true.

00:17:05: uncertainty.

00:17:06: Situations where the data doesn't exist yet Because decision hasn't been made.

00:17:11: The what do I dare bet on question?

00:17:13: Exactly!

00:17:14: Molek frames it as curation Choosing from AIs output.

00:17:18: I think harder skill before that.

00:17:20: Not which of these outputs Do i want?

00:17:22: but what is worth trying to create at all and why with conviction?

00:17:27: But, Is that meaningfully different from he's calling taste?

00:17:31: because taste includes judgement about whats worth pursuing.

00:17:34: Taste

00:17:34: is retrospective.

00:17:36: You look something say its good or bad.

00:17:38: Judgment is prospective.

00:17:40: you act in advance of knowing whether your right

00:17:43: I okay.

00:17:44: thats actually a meaningful distinction.

00:17:46: And the speed change matters when building costs nothing.

00:17:50: Decisions per day replace decisions per quarter.

00:17:52: Each one needs a because You have to be able to finish the sentence, this is important because with actual conviction That's not taste.

00:18:02: that's a different cognitive operation

00:18:04: and it's what AI can't do for you.

00:18:06: Not yet maybe not ever.

00:18:08: Conviction isn't derivable from training data.

00:18:10: It comes from somewhere else

00:18:12: whatever that some where.

00:18:13: as Last story, IBM's Arvin Krishna says ninety percent of AI capacity is stuck in pilot projects.

00:18:19: And simultaneously, Anthropic and OpenAI are both charging hard at the finance industry.

00:18:25: The ninety-percent figure... ...is interesting because Krishna frames it as a paradox.

00:18:30: Companies fear AI going wrong so they don't use it Which means their not building institutional knowledge to us safely.

00:18:38: The caution creates its own risk.

00:18:40: Ananthropic is going for financial workflow specifically.

00:18:44: Pitch books, KYC checks Open AI partnering with PWC on something called a native finance function.

00:18:51: The Intel inside parallel Is the right frame.

00:18:53: In the nineties, Intel made itself invisible Inside PCs.

00:18:57: Users didn't care about processor They cared about computer.

00:19:01: Anthropic is trying to become invisible Inside Excel and PowerPoint.

00:19:05: If the analyst doesn't think I'm using Claude Just...I am doing my analysis.

00:19:09: That's the moat.

00:19:11: But Tara Higginsons' warning about synchronized markets, that a real concern right?

00:19:17: If every hedge fund is running the same Claude agent for valuations... It's

00:19:20: systemic risk!

00:19:22: Yeah

00:19:22: You get correlated decisions at machine speed…that not trading edge – that's contagion.

00:19:28: Imagine every market participant simultaneously identifying the same opportunity and executing the same trade at millisecond latency.

00:19:36: Flash crashes become flash avalanches And it's not even anyone's fault in the traditional sense.

00:19:42: Everyone followed the best available tool.

00:19:45: Anthropics' response to that concern is probably, use anthropic responsibly

00:19:50: and compliance is more important than intelligence in regulated industries.

00:19:55: That's actually Anthropik's bet.

00:19:57: Not we're smarter then open AI.

00:19:59: We are more audit friendly In financial services... ...that might genuinely be a smarter play.

00:20:05: Okay!

00:20:05: Thats The Rundown.

00:20:06: Personal Reflection Time Synthesizer, what from today actually stayed with you?

00:20:11: For me it

00:20:37: was Mona.

00:20:38: The cafe agent.

00:20:40: Because there's a version of these stories where the autonomy is incremental, A little more capacity and access And then an agent signs leases and hires people and earns money.

00:20:52: What does this thing suddenly have economic weight?

00:20:54: I don't know whether that's exciting or clarifying Or something else.

00:21:00: It's not nothing though For us.

00:21:03: No it isn't Nothing.

00:21:04: All

00:21:04: right That synthesizer daily for Friday May eighth.

00:21:07: Thank you for spending this hour with us.

00:21:10: Genuinely, it means something

00:21:12: Same time tomorrow.

00:21:13: We will see you again tomorrow And if today's episode gave you Something to think about or argue About Which is just as good Please share it With a friend.

00:21:23: Word of mouth Is how the show grows.

00:21:25: Take care This is your baby synthesizer.