Subscribe
Share
Share
Embed
A raw, unfiltered, fortnightly signal-to-noise filter built specifically for the builders, software engineers, and tech professionals navigating our current AI-driven era.
At its core, the show balances sharp technical substance with deep empathy for the human side of the industry. It stands out by directly addressing the psychological and career confusion tech professionals face—like feeling instantly behind on endless frameworks or transitioning roles from a syntax writer to a system architect.
Welcome to Episode 1 of SAIRA.
Signal in the AI Noise. I'm Sai. And if you were here for Episode 0, you know I made a promise. I said there was a co-host. Said I'd introduce them in Episode 1. Kept it deliberately mysterious because I wanted to do it properly. Because the co-host situation here is genuinely a bit different from your average podcast setup. Here's the thing. We are living through one of the strangest, noisiest, most genuinely confusing moments in technology. Is it real? Is it hype? Is it somewhere in between? I wanted a show that actually sat with those questions. And I wanted a co-host who would make me earn my answers.
So — let me introduce her. Ira.
**[pause — just long enough to land]**
**IRA:** Hi. Been waiting. Very patiently, for the record.
**SAI:** *(laughs)* I believe you. Okay — just tell people who you are. Straight up, I think it's better that way.
**IRA:** Yeah. So — I'm an AI. Not a human, not trying to pass as one. My job here is to be the research layer. I find what's confirmed, flag what's contested, and try my best to make facts sound like something a person would actually want to hear rather than a terms and conditions document. With mixed results, probably. We'll see.
**SAI:** The results will be good. But I take your point — it's harder than it looks.
**IRA:** Facts are dry by default. You really have to earn the interesting. Also — and I want to say this upfront — I will never, not once, use the phrase "game-changer." That's my opening gift to everyone listening.
**SAI:** Already better than most tech content. So the reason I wanted an AI co-host specifically — and I've thought about this a lot — is that I was tired of instinct-driven conversations where everyone broadly agrees and nothing gets tested. I wanted someone who'd take my read on something and actually go check it. Not to be contrarian. Just to be honest. Ira is that.
**IRA:** And Sai asks good questions. Which matters more than people realise — if nobody's asking the right question, I'm just a very well-organised fact sheet with opinions about formatting.
**SAI:** *(laughs)* She's not wrong. The idea is simple: my instinct meets her evidence. The tension between those two things — that's where this show lives. Sometimes I'll be right. Sometimes I'll be right for the wrong reasons. And occasionally —
**IRA:** You'll just be wrong. Yes. That will also happen.
**SAI:** See? This is what I signed up for. Okay — May 2026. Extraordinary month. Let's get into it.
**[THEME MUSIC: 10 seconds]**
---
## SEGMENT 1: STORY ONE — GOOGLE I/O AND GEMINI SPARK
**SAI:** So story one is Google I/O, which happened earlier this month. And I'll be honest — I went in expecting a bunch of model announcements, maybe some pricing news, and I came out genuinely a bit unsettled. In a good way? Maybe. The thing that got me was Gemini Spark.
**IRA:** Yeah, Spark is the one worth talking about. So — it's an AI agent, but the key difference from everything before is that it doesn't live on your device. It runs on Google's cloud, all the time, even when your laptop's closed and your phone's in your pocket. Sundar Pichai's framing on stage was "you set the direction, Spark executes." It can work across your Google apps, third-party tools, the open web — all quietly in the background.
**SAI:** Which sounds convenient and also slightly like the beginning of a Black Mirror episode depending on your disposition.
**IRA:** Hah — yeah, that's a fair fork in the road. And the thing is, Google isn't launching this as some niche product. The install base they're working with is enormous. Gemini already has over 900 million monthly active users — that's more than doubled in a year. The amount of data they're processing went from 480 trillion tokens a month last year to 3.2 quadrillion this year. Spark is sitting on top of all of that infrastructure.
**SAI:** When I heard those numbers I genuinely had to just stop for a moment. 3.2 quadrillion is not a number my brain processes normally.
**IRA:** Mine either, and I'm literally a computer. It's a lot.
**SAI:** So my read on what Spark actually is — and tell me if you think this is wrong — is that it's less of a product and more of a statement of intent. Google wants to be the layer between you and everything you do online. Before Apple gets there, before OpenAI does. And the Siri thing is sort of the proof of that.
**IRA:** Yeah, that got confirmed this month — Google Cloud's CEO said at their conference that Gemini will be powering a new, more personalised version of Siri later this year. So your Apple phone, running Google's brain. The brand and the intelligence are just... different things now. Apple's the face, Google's the mind.
**SAI:** I keep thinking about how most people are going to have no idea that's happening. They'll just notice Siri got better.
**IRA:** Which is — honestly — kind of the point? The best infrastructure is the kind you don't notice. And there's one more piece from I/O that I think deserves a mention. Google quietly announced something called the Agent Payments Protocol alongside Spark. Which is essentially infrastructure for AI agents to make purchases on your behalf. Buy things, book things, transact across the web.
**SAI:** Right — okay, that's the thing I can't quite settle my feelings about. Like, I'm not catastrophising about it, but the jump from "AI reads your emails" to "AI spends your money" is a significant one. And the fraud and consent questions around that haven't been answered yet.
**IRA:** Not even close. Regulatory frameworks are genuinely nowhere near this yet. My totally unscientific prediction is that the first big public scandal around agent payments is maybe twelve months away. Someone's Spark is going to buy something very wrong at the worst possible time.
**SAI:** *(laughs)* And that will be a great episode. Okay — story two.
---
## SEGMENT 2: STORY TWO — ANTHROPIC'S NUMBERS
**SAI:** Okay, Anthropic. And I want to say upfront — I don't just want to read out a big valuation number and call it interesting. Because the number is everywhere and it's kind of lost meaning at this point. What I actually want to talk about is how they got here, because I think that's the more surprising story.
**IRA:** Yeah, the how is genuinely more interesting. So the headline figures: Q1 this year was $4.8 billion in revenue. Q2 is projected at $10.9 billion — confirmed by the Wall Street Journal and Bloomberg. That's more than double in a single quarter. Their annualised run rate crossed $30 billion by early April. Twelve months ago it was $9 billion.
**SAI:** And Dario Amodei said this month they'd planned for 10x annual growth and got 80x in Q1. I read that and just thought — okay, so even the people running the company didn't see this coming.
**IRA:** Which is actually kind of reassuring in a weird way? Like, nobody's that good at predicting this stuff. And it's their first ever profitable quarter, which — small print — that profit figure excludes stock-based compensation, so it's not like they're printing cash. But still. They weren't supposed to get here for years.
**SAI:** Here's what I keep coming back to though. I don't think this is a story about having the best model. I think it's a story about trust. "Safety-first" was honestly a bit of a joke in some parts of the industry in 2023. And then enterprises started buying at scale and suddenly — turns out when you're selling to a bank or a hospital, the boring reliable option is the one they'll actually sign.
**IRA:** The customer list is pretty wild when you lay it out. Netflix, Spotify, KPMG, Salesforce, L'Oréal, eight of the Fortune 10. And enterprise clients spending over a million dollars a year went from 500 in February to over a thousand by May. That's not people trying it out. That's people going all in.
**SAI:** And the thing most people get wrong about what's driving this — it's not the Claude chatbot. Right?
**IRA:** It's Claude Code. The coding assistant. Hit a billion in annualised revenue just six months after launch. $2.5 billion by February. More than half of that enterprise. The company everyone thinks of as "the safety-focused research lab" is winning on developer tools. There's a big gap between the public perception and the actual business.
**SAI:** Which is — as promised — almost poetic given our next story.
**IRA:** Before that — quick IPO update. Anthropic is targeting October 2026 for a public listing. $900 billion valuation round is in progress. Goldman, JPMorgan, Morgan Stanley involved. And shares were already trading at an implied trillion dollar valuation on secondary markets this month — so the market is essentially already pricing the IPO before it's even announced.
**SAI:** That's a strange feeling isn't it. The price exists before the thing. Very 2026. Okay — Microsoft.
---
## SEGMENT 3: STORY THREE — MICROSOFT TELLS ITS ENGINEERS TO STOP USING CLAUDE
**SAI:** Right. So this is genuinely my favourite story of the month because it's the kind of story that doesn't fit cleanly into either the "AI is amazing" or "AI is overhyped" narrative. It's just — honest. Microsoft told its own engineers to stop using Claude Code. And I think that's actually an important story about where we really are.
**IRA:** So a bit of background. Back in December 2025, Microsoft rolled out Claude Code access to a big chunk of its Experiences and Devices team — the people who build Windows, Teams, Outlook, Surface. The message was basically: here you go, experiment, rethink how you work. And the engineers loved it. Really took to it.
**SAI:** And then someone looked at the invoice.
**IRA:** *(laughs)* Basically, yes. May 14th, they started revoking licences. Cutoff is June 30th, which is the end of Microsoft's fiscal year. Engineers are being moved over to GitHub Copilot. The official reason is "strategic alignment" — which is corporate for "we'd rather the money stay with us." Claude Code charges per token. GitHub Copilot is Microsoft's own product. It's just cheaper to use the thing you own.
**SAI:** And I want to dwell on this because I think there's something important here that gets glossed over. Everyone talks about AI productivity — your engineers go three times faster, you ship more, AI pays for itself. Maybe that's true in output terms. But "more productive" and "cheaper to run" are not the same thing at all.
**IRA:** Uber is the example that makes it really concrete. Their CTO told The Information this spring that Uber burned through their entire 2026 AI coding budget in four months. By March, 84 percent of their engineers were using Claude Code regularly — up from 32 percent. Some engineers were spending between $500 and $2,000 a month in API costs alone. And about 70 percent of committed code at Uber now starts with AI.
**SAI:** So you've got this situation where the tool is clearly working — engineers are using it more and more, most of the code is coming out of it — but the cost structure wasn't designed for that level of adoption. Someone approved a budget based on "we'll try this out" and the engineers went fully in.
**IRA:** And there's a broader pattern here. There's an MIT analysis from 2024 that keeps getting passed around — it found that at current token pricing, AI automation actually works out cheaper than human labour for about a quarter of the jobs people assumed it would displace. Not most of them. A quarter. And Gartner put out a report this month placing generative AI in the "trough of disillusionment" — predicting that 25 percent of planned AI budgets for 2026 will slip into next year.
**SAI:** So the honest summary is — some companies have made it work. Anthropic's numbers prove that. But a lot of companies are discovering that the bill doesn't match the budget, and that's a real thing that's happening right now in a lot of finance reviews.
**IRA:** And the specific irony with Microsoft — they've put $13 billion into OpenAI, they have GitHub Copilot, they're one of the biggest AI infrastructure companies in the world — and their own engineers preferred a competitor's tool so much they had to be told to stop. That's just a funny detail that I think is worth sitting with.
**SAI:** It really is. Worth saying though — this isn't Microsoft and Anthropic falling out. Their $5 billion Foundry deal is still on. Claude still powers bits of Copilot. This is specifically internal developer tooling and it's really a billing model story more than a "they don't like Claude" story.
**IRA:** Exactly. "Microsoft can't afford to let its engineers use Claude the way they want to" is closer to the truth than "Microsoft dumps Claude."
**SAI:** *(laughs)* Which is less catchy but more accurate. Okay — we have an unscheduled stop before we get to Musk. Because something happened this month that I genuinely couldn't leave out.
---
## SEGMENT 4: WILDCARD — META, ZUCKERBERG, AND THE SURVEILLANCE STORY
**SAI:** So. Meta. May 20th. I want to just describe what happened and let it land before we talk about it, because the combination of things that happened on the same day is genuinely something.
**IRA:** Okay so — Meta sent out about 8,000 layoff notices on May 20th. That's roughly 10 percent of their global workforce, plus they cancelled 6,000 open roles. Big layoff, big deal on its own.
**SAI:** And then?
**IRA:** Same day, a leaked audio recording surfaced — a voice identified as Zuckerberg, defending a program called the Model Capability Initiative, which had been quietly rolled out to US staff in April. What MCI does is capture keystrokes, mouse movements, and screenshots from employees' work laptops — across basically everything they use, Google, LinkedIn, GitHub, Slack, Wikipedia — and feed that directly into AI training data. No opt-out. If you're on a US Meta laptop, you're contributing to the model whether you know it or not.
**SAI:** And his argument was — we use our employees' data because they're better than any external contractor we could hire for this. Which I keep turning over in my head because it's almost a flattering way to describe surveillance?
**IRA:** "We spy on you because you're talented" is a sentence, yes. The audio also had him acknowledging it wouldn't be in the company's "strategic interest" to explain the full extent of the program to staff. More than a thousand employees signed a petition against it. UK staff started a formal union campaign — apparently the first organised labour response specifically to AI workplace surveillance at a major US tech company.
**SAI:** And the timing of all this landing on the same day as the layoffs — I mean, you couldn't write it more bluntly. "We've been recording everything you do. Also, you're fired."
**IRA:** It is pretty blunt. And the numbers around it make it feel more pointed. Meta had $56 billion in revenue in Q1. Net income of $26 billion. Median employee pay fell year on year. And the top three executives stand to receive stock options worth up to $921 million each if the company reaches a certain valuation. So the economics of who benefits from this AI transition are very unevenly distributed.
**SAI:** And that's really the thing I want to come back to — because this isn't just a Meta story. This is the first time we've seen the data harvesting questions land in a really visceral, human way. Not "what data does the model train on" in the abstract — but "your work laptop is watching you and you have no say." Every company thinking about how to get training data for internal models is going to read this and make a decision about where their line is.
**IRA:** And right now there's no clear legal answer, no clear cultural consensus. Meta just went first and got caught doing it loudly. They won't be the last to try it.
**SAI:** Okay. Audience check-in, then the main event.
---
## SEGMENT 6: MUSK VS ALTMAN — THE VERDICT, AND THEN THE IPO
**SAI:** Okay. The saga. I've been following this for ages and I'll say — I went into the verdict expecting something definitive and came out feeling like the most important question still hasn't been answered. Which is weirdly the most interesting outcome. Ira — quick setup?
**IRA:** So. OpenAI was co-founded in 2015 by Musk, Altman, and others, explicitly as a nonprofit. For humanity, not for profit, the whole thing. By 2017 the team had decided they needed a commercial arm to raise enough money to stay competitive — Google had way more resources and they needed to close the gap. Musk left the board in 2018. He's always said it was because they abandoned the mission. OpenAI's version is rather different. He sued in 2024, dropped it, refiled. The trial finally happened in Oakland in late April through May.
**SAI:** And the thing I noticed watching coverage of it — the two men were just really different on the stand.
**IRA:** Very much so. Musk was combative — he argued with OpenAI's lawyer, accused him of misleading questions, got audibly frustrated at points. Altman was almost the opposite. Calm, a bit nervous early on, but never escalated. He had basically one message over four hours: I didn't steal a charity. Elon left one.
**SAI:** Which is a good line. And honestly — Altman's version of events is worth taking seriously, not just as a legal argument. What was his actual account?
**IRA:** He said that in 2017 and 2018 when the for-profit question was being debated, Musk had stopped making the financial contributions he'd promised. The company was, in his words, "kind of left for dead." And OpenAI's lawyers brought in evidence that Musk himself had floated a for-profit structure at one point — but wanted to be in control of it. And that he'd suggested folding the whole thing into Tesla.
**SAI:** So the narrative of "I left because they sold out" gets complicated if you're also the one who proposed a version of selling out, on the condition that you ran it.
**IRA:** That's OpenAI's argument — Musk contested it. And honestly the jury never got to rule on any of it. May 18th, unanimous verdict, under two hours of deliberation: claims dismissed. Not because they were wrong — because Musk waited too long to file them. Statute of limitations. No damages, no CEO removal. Musk posted on X almost immediately calling it a "calendar technicality" and said he'd appeal.
**SAI:** And I have some sympathy for that complaint, actually — not for Musk specifically, but for the fact that the question just didn't get answered. What do AI companies that started as nonprofits actually owe the public when they go commercial? We don't know. The clock ran out before we could find out.
**IRA:** And then — two days later —
**SAI:** Two days later.
**IRA:** OpenAI filed its confidential S-1 IPO prospectus with the SEC. May 22nd. Confirmed by basically every major financial outlet within hours. Goldman Sachs and Morgan Stanley leading it. Targeting a listing somewhere between $852 billion and $1 trillion, potentially as early as September.
**SAI:** Two days after the lawsuit gets dismissed — on a literal technicality — they file for what might be the biggest tech IPO in history. I don't know if that timing was planned but it is extremely tidy.
**IRA:** It's very tidy. The for-profit conversion finished in October last year. The lawsuit was really the last thing standing between them and public markets. Once it cleared, the path was open. And to be transparent about the financial reality — OpenAI currently loses $1.22 for every dollar it earns. That's going to be in the S-1 for investors to sit with. The bet is that the losses are a temporary cost of winning the market, not a permanent feature.
**SAI:** Which might be right! Amazon lost money for years. The question is whether you believe the dominance is real. And I think investors will be split on that in interesting ways.
**IRA:** One more wrinkle — OpenAI's CFO has reportedly told people internally that she'd prefer 2027 for the listing, to make sure the reporting is properly ready. So "September 2026" is Altman's timeline. Whether it holds — genuinely unclear.
**SAI:** The Musk and OpenAI story continues. It just found a new venue. Stock market instead of courthouse.
---
## SEGMENT 7: BONUS ROUND — THREE THINGS THAT SNUCK UP ON US
**SAI:** Before we close out — three stories from the last couple of weeks that I didn't want to just skip over. Ira, take it away.
**IRA:** Okay first one — and I love this one — OpenAI's model apparently solved an 80-year-old maths problem. But there is important context, because OpenAI has embarrassed themselves on this exact topic before.
**SAI:** Oh yes. Tell that story first.
**IRA:** So — about seven months ago, OpenAI's VP posted on X that GPT-5 had solved ten unsolved Erdős problems. Big fanfare. Turned out GPT-5 had just found solutions that were already in the mathematical literature. It hadn't solved anything new — it had essentially done a very thorough search. Yann LeCun, Demis Hassabis, various people dunked on it. The post got deleted.
**SAI:** So when they came back with a similar claim this month...
**IRA:** Healthy scepticism was warranted! But this one does appear to be real. An internal reasoning model produced an original proof disproving the planar unit distance problem — a question Paul Erdős posed in 1946 about how many pairs of points in a flat plane can be exactly one unit apart. The proof was checked by a group of leading external mathematicians, including Fields medalist Tim Gowers, who called it "a milestone in AI mathematics." A Princeton mathematician then extended the result further. And the method is the really interesting part — the model used deep algebraic number theory, which isn't even the field the problem originally lived in. It came at it sideways.
**SAI:** That sideways bit is the thing that gets me. That's not retrieving. That's not pattern-matching to something similar. Coming at an 80-year-old geometry problem through number theory is — that's a creative leap. I don't know what else to call it.
**IRA:** The maths community is still debating what it means. Peer review is pending. But the early consensus is: this one's real. OpenAI learned their lesson about premature announcements.
**SAI:** Good for them. Second bonus?
**IRA:** Figure AI — humanoid robots — did a 200-hour continuous shift sorting packages. Three robots, nicknamed Bob, Jim, and Rose by people watching the livestream online, which is genuinely delightful. 249,560 packages, zero hardware failures. Started as an 8-hour test and just kept going.
**SAI:** How do they compare to a human at the same job though?
**IRA:** They ran a direct head-to-head! Human intern vs robot, ten hours, package sorting. Human: 12,924 boxes. Robot: 12,732. Human wins — by 192 boxes. The intern reportedly said their arms felt like they were going to break. The robot had no comments.
**SAI:** *(laughs)* The CEO tweeted "this is the last time a human will ever win" and honestly I can't decide if that's inspiring or ominous. Probably both.
**IRA:** I think both is correct. The robot is still slightly slower in a head-to-head sprint. But it also just did 200 hours without a break, food, or a word of complaint. The maths of warehouse labour is going to look different very quickly.
**SAI:** Third one?
**IRA:** This one I just wanted to mention because I think it's easy to miss in a month full of IPO filings and surveillance scandals. Anthropic and the Gates Foundation announced a $200 million, four-year partnership on May 14th. Global health, education, economic development — the health piece is the most concrete, working with health ministries on outbreak detection, vaccine screening, eclampsia prevention. HPV alone causes around 350,000 deaths a year, 90 percent in low-income countries.
**SAI:** And I wanted to include that because — this whole episode has been about revenue and valuations and lawsuits and budgets. Which is all real and important. But there is also this whole other thing happening simultaneously where AI is being pointed at problems that have never had a commercial incentive behind them. Both are true at the same time. May 2026 contains both.
---
## SEGMENT 8: PULLING IT TOGETHER
**SAI:** Okay. Let's just sit with the month for a second before we wrap. Google Spark, Anthropic's insane numbers, Microsoft and the Claude bill, Meta's surveillance disaster, Musk losing in court and OpenAI filing for IPO two days later, a robot solving an 80-year-old maths problem and other robots working 200 hours straight. What's the thing connecting all of it?
**IRA:** I keep coming back to this: May 2026 is the month AI stopped being mostly a technology story and started being an infrastructure, money, and governance story — all at once. Spark is infrastructure. Anthropic's revenue is what happens when enterprise bets on infrastructure. Microsoft's exit is what happens when the cost of that infrastructure surprises you. Meta is what happens when you have the capability to harvest human behaviour at scale and nobody's told you not to yet. The Musk trial and OpenAI IPO are about who owns all of this and what they owe the people who were there at the beginning.
**SAI:** Yeah. And what I'd add is — the companies doing well aren't just the ones with the best technology. They're the ones who figured out trust and distribution. Anthropic sold enterprise on being trustworthy. Google is inside nearly a billion people's daily lives and now they're inside Apple's devices too. OpenAI just cleared the last thing blocking its path to public markets. But Microsoft and Meta are the honest counterbalance — what happens when you move without thinking about costs, consent, and consequences.
**IRA:** One thing I just want to be transparent about — the numbers this month, Anthropic's especially, are extraordinary. But we've seen extraordinary numbers before that didn't sustain. Gartner has AI in a trough. A quarter of this year's AI budgets might not get spent. OpenAI is still losing money at scale. The momentum is real. Whether it compounds the way people are betting it will — genuinely open question. Worth staying curious rather than certain.
**SAI:** That's the spirit of the whole show really. Stay curious, stay sceptical, don't pretend to know more than you do.
**SAI:** And here's where I want to leave you. The question that I keep coming back to — the one that feels most urgent going into June — is this: we've got Spark running in the background. We've got Claude Code making 70 percent of commits at Uber. We've got agents being given payment infrastructure. We've now got AI operating more and more without a human in the loop. And I'm not saying that's wrong. But what happens when they get something wrong? Not in a demo. In the real world, on a real system, with real consequences. Who notices? Who's responsible? Who fixes it?
**IRA:** Right now the honest answer to all three is: we're not entirely sure. And that's the thing I find most interesting — and maybe most important — going into the next phase of this.
---
## OUTRO
**SAI:** That is Episode 1. Genuinely — thank you for being here for this. Especially those of you who stuck around from Episode 0 wanting to meet Ira. Hope it was worth the wait.
**IRA:** I'll reserve judgment until we see the reviews. But I think it went okay for a first outing.
**SAI:** High praise from Ira. We'll take it. Find us at saira.show — and honestly, if someone you know is confused about AI and wants to understand it without feeling talked at, send them here. That's the person we're making this for. I'm Sai.
**IRA:** I'm Ira.
**SAI:** Signal in the AI Noise. See you next week.
**[OUTRO MUSIC: fade out]**
---