The guy running Microsoft AI just signed his name to a sentence — superintelligence is near, and don’t worry, it won’t take your job. This is the Tech Podcast Podcast. Today: Mustafa Suleyman on the record, up against a BAML engineering episode about debugging agents that have already gone off the rails. One guy says the future’s almost here. The other guys are asking who pays for the three days you spent figuring out why your agent broke. Love that split. So let’s start at the top of the food chain. A sitting CEO making a timeline-ish claim — that deserves harder questions than anything we’ve heard all week. Suleyman is world-class at sounding candid while telling you nothing. So did the host actually pry anything loose, or did the farm backdrop make everybody cozy? The bit that caught me was his approach to training new models. That’s the line that sounds operational. Did it stay operational? Exactly. That’s the tell. “Approach to training” sounds specific and reveals nothing structural. If he didn’t say what data, what scale, what failure rate — it’s the Copilot PR move all over again. And “won’t take your job” isn’t neutral — it’s the most prominent answer yet to the question we’ve been asking. After Kedrosky’s meteor read, and Foday saying tokens already exceed salaries, now the CEO says relax. Relax — while we’ve spent the week watching token bills blow past budgets. Did the host put one ounce of that cost-side pressure to him, or just nod through the comfortable part? Which is why the BAML episode lands so well right next to it. Suleyman talks superintelligence in the abstract, while these engineers are asking who owns the P&L when an autonomous agent fails non-deterministically. This is the one for me. If you can’t reproduce the failure three days later, that’s a cost multiplier nobody’s pricing in. The liability question, but in production form. And it answers Kilpatrick’s “reset your ambition every two-to-three months” in the least glamorous way possible — your ambition resets when the agent breaks and observability can’t tell you why. Right, that’s what the heuristic looks like on the ground — a debugger, not a manifesto. Queue the BAML episode first. The CEO can wait. Daily Guardian writes:
We covered everything from Mustafa’s approach to training new models to his criticisms of Anthropic talking about Claude as though it is conscious. Of course, we also talked about Microsoft’s relationship with OpenAI, how Mustafa is thinking about all the negative polling and political pushback around AI right now, and whether any of the consumer products are good enough to overcome it.
So the Microsoft AI CEO sits down on Decoder and leads with “superintelligence is near, but it won’t take your job.” That may be the safest sentence a man in his position can say. It’s the fourth jobs take we’ve gotten this week — and the first from someone actually running a frontier lab. Suleyman co-founded DeepMind, so when he attaches a timeline to superintelligence, that’s worth pressing on. Did he attach a timeline, though? “Near” is not a number. We had SpaceX putting a fixed price down earlier this week — Suleyman’s giving conviction with nothing you can hold him to. The tell for me is the phrase the host flagged — Mustafa’s approach to training new models. That’s where you’d get something structural. Did Patel actually pull any of that out, or did it stay at altitude? And the juicy bit nobody’s talking about — Suleyman taking a shot at Anthropic for talking about Claude like it’s conscious. That’s a frontier lab CEO calling out a rival’s marketing on the record. Which is convenient. Microsoft’s man draws the sober line on consciousness while Anthropic does the mysticism — and somehow both serve the brand. I want him pressed on cost, not vibes. The token bills don’t care how we’re supposed to feel. From BAML Podcast:
In this episode, we will dive into AI agent observability and answer a question every production engineer eventually faces: how do you diagnose why an autonomous agent went off the rails three days ago? When you are dealing with non-deterministic tools, old-school debugging habits like inserting print("here") statements fails to scale.
Right after a Microsoft CEO tells us superintelligence is near and won’t take your job — here’s the BAML crew asking how you even figure out why your agent went sideways three days ago. That’s the gap, in two episodes back to back. And it’s a real cost question. If a non-deterministic failure takes three days to diagnose, one bug ticket turns into engineer-weeks of P&L on top of the token bill we spent all week tracking. My favorite line in the writeup: print “here” fails to scale. Every production engineer who’s ever debugged an agent just nodded so hard they hurt themselves. What sells me is the replay piece — reconstructing the agent’s exact decision tree. That’s the difference between guessing and actually owning the failure. Suleyman talks about agents working for you; nobody up there talks about who pays when they don’t. It loops back to that liability question — when an autonomous process goes off the rails, who owns it? Turns out step one is just being able to see what it did. Wild how far down the stack the honest answers live. Got feedback, a story idea, or a correction for us? Send it to techpodcastpodcast at lantern podcasts dot com. We read what you send, and it helps us make the show sharper.
You’ll find links to every story we mentioned today in the show notes, so if something stuck with you, that’s the place to dig in a little further.
That’s Tech Podcast Podcast for this Tuesday, June 9th. Thanks for listening. This is a Lantern Podcast.