Latest / The AI Practitioner Brief / Anthropic's Cuts, DeepSeek's Cache, and CoreWeave Debt — October 19, 2026

Transcript

Anthropic's Cuts, DeepSeek's Cache, and CoreWeave Debt — October 19, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_01: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_01: the podcast from AIPractitioner.substack.com,
• 0:11
  SPEAKER_01: where I turn the latest AI in tech news into practical insights
• 0:14
  SPEAKER_01: for AI practitioners.
• 0:16
  SPEAKER_01: I'm Lina Faik,
• 0:17
  SPEAKER_01: an ML engineer,
• 0:18
  SPEAKER_01: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:23
  SPEAKER_01: It's Monday, October 19th.
• 0:25
  SPEAKER_00: Monday already.
• 0:26
  SPEAKER_00: Let's get into it.
• 0:27
  SPEAKER_01: With me is Jimmy, my co-host on the brief, who asks the questions I'd otherwise get past.
• 0:32
  SPEAKER_00: That's the job description, yes.
• 0:34
  SPEAKER_01: Anthropic published its own hiring data last week.
• 0:37
  SPEAKER_01: Junior technical hires dropped from 27% of new hires to 9% in about a year.
• 0:43
  SPEAKER_01: On the same day,
• 0:44
  SPEAKER_01: bond traders were pricing one of the biggest AI data center builders as more likely
• 0:48
  SPEAKER_01: than not to default.
• 0:50
  SPEAKER_00: So the company selling the agents is also
• 0:52
  SPEAKER_00: the one proving what they cost somebody else's career.
• 0:55
  SPEAKER_01: That's roughly it.
• 0:56
  SPEAKER_01: Headlines first, then three bigger stories.
• 0:59
  SPEAKER_01: Anthropics' own numbers show what's happening to junior jobs.
• 1:02
  SPEAKER_01: DeepSeek just shrank its flagship model's memory footprint by a lot.
• 1:07
  SPEAKER_01: And bond traders stopped trusting the biggest AI landlord's debt.
• 1:10
  SPEAKER_00: Let's start with what's actually new today.
• 1:13
  SPEAKER_01: Replit, a browser-based coding platform,
• 1:15
  SPEAKER_01: switched the default AI engine behind its coding assistant from Anthropics Cloud
• 1:20
  SPEAKER_01: to OpenAI's models.
• 1:21
  SPEAKER_01: The company says it came down to cost, performance, and partnership considerations.
• 1:26
  SPEAKER_00: Can you still pick Cloud if you want it?
• 1:28
  SPEAKER_01: Yes.
• 1:29
  SPEAKER_01: Claude and other models stay available as options.
• 1:32
  SPEAKER_01: OpenAI just became the default.
• 1:34
  SPEAKER_00: What's the other AI story today?
• 1:36
  SPEAKER_01: Even the companies building AI tools admit their own pricing confuses people.
• 1:40
  SPEAKER_01: A growth lead at Lovable,
• 1:42
  SPEAKER_01: a startup that builds apps from plain language prompts,
• 1:45
  SPEAKER_01: says a $25 plan buys you 100 credits,
• 1:48
  SPEAKER_01: but you don't know what those credits actually do until you start building.
• 1:52
  SPEAKER_00: That's not a pricing plan, that's a lottery ticket.
• 1:54
  SPEAKER_01: It gets worse with general-purpose AI,
• 1:56
  SPEAKER_01: where the same request can cost wildly different amounts depending on how many times
• 2:01
  SPEAKER_01: an agent retries itself before finishing.
• 2:03
  SPEAKER_00: Is there a fix anyone's landed on?
• 2:05
  SPEAKER_01: Not yet.
• 2:05
  SPEAKER_01: Not the way audio tools settled on selling minutes.
• 2:08
  SPEAKER_01: That's the AI side.
• 2:10
  SPEAKER_01: Elsewhere in Tech Today,
• 2:11
  SPEAKER_01: reports describe a future, pricier version of Apple's HomeView Smart Display.
• 2:16
  SPEAKER_00: The one with the robot arm?
• 2:17
  SPEAKER_01: Right.
• 2:17
  SPEAKER_01: The current home view costs $299 and is mostly voice controlled.
• 2:22
  SPEAKER_01: The version in these reports adds a motorized arm that physically turns the screen
• 2:27
  SPEAKER_01: toward whoever's talking.
• 2:28
  SPEAKER_00: So the current model's already behind what they're planning.
• 2:31
  SPEAKER_01: That's the read.
• 2:32
  SPEAKER_01: Meanwhile,
• 2:32
  SPEAKER_01: Datadog,
• 2:33
  SPEAKER_01: which makes monitoring software for engineering teams,
• 2:36
  SPEAKER_01: introduced Scout,
• 2:37
  SPEAKER_01: a specialized model built just
• 2:38
  SPEAKER_01: for investigating production incidents inside its Bits AI assistant.
• 2:42
  SPEAKER_00: Well, what's the pitch on that one?
• 2:44
  SPEAKER_01: The company says it matches general frontier models at three
• 2:47
  SPEAKER_01: to eight times lower cost,
• 2:49
  SPEAKER_01: according to its own testing,
• 2:51
  SPEAKER_01: trained on real incident replays instead of general web text.
• 2:54
  SPEAKER_00: So it's not about being smarter, it's about being cheap enough to run on everything.
• 2:58
  SPEAKER_01: That's the pitch, yes.
• 3:01
  SPEAKER_00: What about the database story?
• 3:02
  SPEAKER_01: Cloudflare,
• 3:03
  SPEAKER_01: which handles traffic for a large share of the internet,
• 3:06
  SPEAKER_01: detailed how its D1 platform,
• 3:08
  SPEAKER_01: built on EscoLite,
• 3:09
  SPEAKER_01: a lightweight embedded database, scales to millions of separate customer databases.
• 3:14
  SPEAKER_01: The trick is keeping slow administrative work,
• 3:17
  SPEAKER_01: like provisioning new databases, completely separate from live query traffic.
• 3:21
  SPEAKER_00: Why would that matter to anyone outside Cloudflare?
• 3:23
  SPEAKER_01: Because it's the same problem anyone building shared infrastructure for many customers hits.
• 3:29
  SPEAKER_01: Slow background work can't be allowed to add delay to the life path.
• 3:32
  SPEAKER_00: That's five.
• 3:33
  SPEAKER_00: Let's get into the first big one.
• 3:34
  SPEAKER_01: Let's start with Anthropic's own report on its workforce,
• 3:38
  SPEAKER_01: published last Friday, called How Employment is Changing at Anthropic.
• 3:42
  SPEAKER_00: Anthropic studying itself.
• 3:44
  SPEAKER_00: That's an unusual kind of transparency.
• 3:46
  SPEAKER_01: Its internal economics team pulled from hiring records,
• 3:49
  SPEAKER_01: calendar logs, slack activity, and agent usage data going back to early 2025.
• 3:55
  SPEAKER_01: They also ran 17 employee interviews.
• 3:58
  SPEAKER_00: And the headline number?
• 3:59
  SPEAKER_01: Junior technical hires fell from 27% of new hires to 9% in about a year.
• 4:05
  SPEAKER_01: The share of junior staff across technical teams overall dropped from 21% to 14%.
• 4:11
  SPEAKER_00: That's not a dip, that's a collapse.
• 4:13
  SPEAKER_01: The report says it's not layoffs.
• 4:15
  SPEAKER_01: Overall headcount is still growing.
• 4:17
  SPEAKER_01: It's that the junior slice of new hiring shrank fast,
• 4:20
  SPEAKER_01: starting when job postings fell off in December last year.
• 4:23
  SPEAKER_00: Why would that hit juniors harder than anyone else?
• 4:26
  SPEAKER_01: Junior engineers traditionally get the routine work.
• 4:28
  SPEAKER_01: Writing boilerplate, fixing small bugs, writing tests.
• 4:32
  SPEAKER_01: That's exactly the work coding agents are best at right now.
• 4:35
  SPEAKER_00: Is this just an anthropic thing, or is it everywhere?
• 4:38
  SPEAKER_01: There's a broader pattern behind it.
• 4:40
  SPEAKER_01: The report cites separate industry data showing about a third
• 4:43
  SPEAKER_01: of machine learning managers and a fifth
• 4:45
  SPEAKER_01: of software managers say they're hiring fewer people specifically because of AI.
• 4:50
  SPEAKER_00: So Anthropic's not an outlier, just the one willing to publish its own numbers.
• 4:54
  SPEAKER_01: That's fair.
• 4:55
  SPEAKER_01: And the report flags a new problem that comes with it.
• 4:58
  SPEAKER_01: Staff told researchers productivity went up using Claude,
• 5:01
  SPEAKER_01: but reviewing and verifying agent-generated code at scale became its own bottleneck.
• 5:06
  SPEAKER_00: What does that bottleneck actually look like day to day?
• 5:10
  SPEAKER_01: Someone senior has to read every pull request an agent generates at
• 5:14
  SPEAKER_01: a volume no team-sized around junior-level work was ever built to review.
• 5:18
  SPEAKER_00: So the work didn't disappear, it just moved up a level?
• 5:20
  SPEAKER_01: Right, from writing code to policing it.
• 5:23
  SPEAKER_01: And that leaves an open question the report doesn't answer if juniors aren't hired
• 5:27
  SPEAKER_01: to learn on real production code,
• 5:29
  SPEAKER_01: who becomes the senior engineer reviewing agents in 10 years?
• 5:32
  SPEAKER_00: That's a problem with no obvious near-term fix.
• 5:35
  SPEAKER_01: Neantropic doesn't try to solve it in the report either.
• 5:38
  SPEAKER_01: They just name it.
• 5:39
  SPEAKER_00: What do you do with that if you're hiring right now?
• 5:41
  SPEAKER_01: Check whether your own junior pipeline is shrinking for the same reason.
• 5:45
  SPEAKER_01: Not because of budget,
• 5:46
  SPEAKER_01: but because the easy tasks that used to justify a junior hire don't exist anymore.
• 5:51
  SPEAKER_00: Easy to miss until it's already happened.
• 5:53
  SPEAKER_01: And it's happening at the exact moment the agents doing that work are also getting
• 5:57
  SPEAKER_01: a lot cheaper to run.
• 5:58
  SPEAKER_00: Cheaper how?
• 5:59
  SPEAKER_01: DeepSeq released an update to its flagship model called DeepSeq-V4.1-Flash this past week.
• 6:06
  SPEAKER_01: It cuts the model's memory footprint while running to roughly a quarter of what
• 6:10
  SPEAKER_01: the previous version used.
• 6:11
  SPEAKER_01: And the version of that memory stored longer term to about an eighth.
• 6:15
  SPEAKER_00: What is this memory actually doing?
• 6:17
  SPEAKER_01: It's called a KV Cache, short for Key Value Cache.
• 6:20
  SPEAKER_01: It's the short-term memory a model keeps open during a conversation
• 6:24
  SPEAKER_01: so it doesn't recompute everything from scratch on every new word.
• 6:28
  SPEAKER_01: The bigger that memory, the more it costs to serve one long conversation, or thousands at once.
• 6:33
  SPEAKER_00: So smaller cache means cheaper to run.
• 6:35
  SPEAKER_01: Exactly, and it means a company can serve more users at once on the same hardware.
• 6:40
  SPEAKER_01: Which matters a lot for anyone running agents that hold long conversations.
• 6:44
  SPEAKER_00: How'd they actually shrink it?
• 6:45
  SPEAKER_01: Two tricks stacked together.
• 6:47
  SPEAKER_01: First, cross-layer reuse,
• 6:49
  SPEAKER_01: where some layers in the model skip building their own full memory and just borrow
• 6:53
  SPEAKER_01: the one the layer before it already built.
• 6:55
  SPEAKER_00: Doesn't that lose information?
• 6:56
  SPEAKER_01: Not entirely.
• 6:57
  SPEAKER_01: The model has three settings per layer,
• 6:59
  SPEAKER_01: build its own from scratch,
• 7:01
  SPEAKER_01: borrow the memory but double-check it against fresh data,
• 7:04
  SPEAKER_01: or just borrow it outright.
• 7:05
  SPEAKER_01: It picks per layer.
• 7:06
  SPEAKER_00: And the second trick?
• 7:07
  SPEAKER_01: Drinking the number format itself.
• 7:10
  SPEAKER_01: The cache normally stores numbers in a format that takes up a fair amount of space.
• 7:14
  SPEAKER_01: This version stores them in 4 bits instead,
• 7:17
  SPEAKER_01: with a scaling adjustment applied every 16 numbers to limit the precision lost.
• 7:21
  SPEAKER_00: 4 bits sounds like it should break something.
• 7:23
  SPEAKER_01: That's the obvious worry, and the paper behind the model addresses it directly.
• 7:28
  SPEAKER_01: According to its own published results,
• 7:30
  SPEAKER_01: performance came out better than the previous version,
• 7:32
  SPEAKER_01: not worse, despite the smaller cache.
• 7:35
  SPEAKER_00: Better at what, specifically?
• 7:36
  SPEAKER_01: Reasoning, coding, and agent tasks, plus the vision benchmarks.
• 7:40
  SPEAKER_01: The paper doesn't break out how much of that gain comes from
• 7:43
  SPEAKER_01: the cache change versus other improvements in the same release.
• 7:46
  SPEAKER_00: So you can't isolate the one trick that's doing the work?
• 7:49
  SPEAKER_01: Not cleanly, no.
• 7:50
  SPEAKER_01: What's clear is the net effect,
• 7:52
  SPEAKER_01: a model that costs less to serve per conversation and still scores better.
• 7:56
  SPEAKER_00: Who actually feels this, practically?
• 7:58
  SPEAKER_01: Anyone running coding agents or long support sessions at scale.
• 8:02
  SPEAKER_01: Cache size per token is becoming as important as the leaderboard score,
• 8:06
  SPEAKER_01: because it's what limits how many conversations you can serve on one GPU.
• 8:10
  SPEAKER_00: Is that the kind of thing most teams even check before picking a model?
• 8:13
  SPEAKER_01: Most still pick on benchmark score alone.
• 8:15
  SPEAKER_01: That's starting to look like the wrong single number to optimize for.
• 8:19
  SPEAKER_00: And this is happening while everyone's trying to make these models cheaper anyway.
• 8:23
  SPEAKER_01: Which is the part that should worry whoever's betting enormous amounts of debt
• 8:26
  SPEAKER_01: on AI compute demand never slowing down.
• 8:29
  SPEAKER_00: Wait, which debt are we talking about?
• 8:30
  SPEAKER_01: Coreweave,
• 8:31
  SPEAKER_01: a cloud company that rents out Nvidia's AI chips to customers like Meta,
• 8:35
  SPEAKER_01: Microsoft, and OpenAI.
• 8:37
  SPEAKER_01: The New York Times reported last week that bond traders are now pricing its risk
• 8:41
  SPEAKER_01: of default at over 60%.
• 8:43
  SPEAKER_00: 60% is not a rounding error.
• 8:44
  SPEAKER_00: No.
• 8:46
  SPEAKER_01: Its bonds are yielding above 15%,
• 8:48
  SPEAKER_01: which puts them in junk territory,
• 8:50
  SPEAKER_01: the label for debt the market sees as highly risky.
• 8:53
  SPEAKER_00: How does a company end up that exposed?
• 8:55
  SPEAKER_01: Corweave buys the chips,
• 8:56
  SPEAKER_01: signs leases of 15 years or more with data center landlords,
• 9:00
  SPEAKER_01: then releases that capacity to the big AI labs.
• 9:03
  SPEAKER_01: It carries all the costs in between.
• 9:05
  SPEAKER_00: So it's borrowing short-term risk to sell long-term promises.
• 9:09
  SPEAKER_01: That's a sharp way to put it.
• 9:10
  SPEAKER_01: It's carrying $72 billion in debt across 51 data centers.
• 9:15
  SPEAKER_01: Betting that demand for that capacity stays strong for well over a decade.
• 9:19
  SPEAKER_00: What happens if it doesn't?
• 9:20
  SPEAKER_01: Permitting delays are already slowing construction across the sector,
• 9:24
  SPEAKER_01: and CoreWeave's founders have sold $3.3 billion of stock since
• 9:28
  SPEAKER_01: the company went public.
• 9:29
  SPEAKER_00: That second part doesn't inspire confidence.
• 9:31
  SPEAKER_01: It's a signal, not proof by itself.
• 9:34
  SPEAKER_01: The International Monetary Fund flagged this as one of three risks
• 9:37
  SPEAKER_01: that could tighten financial conditions sharply.
• 9:40
  SPEAKER_01: Debt from the handful of giant cloud companies now makes up 10%
• 9:44
  SPEAKER_01: of global investment-grade bond issuance, up from under 1% two years ago.
• 9:48
  SPEAKER_00: That's a fast climb for one sector.
• 9:50
  SPEAKER_01: It is.
• 9:51
  SPEAKER_01: And it's worth separating CoreWeave from Google, Microsoft, and Amazon here.
• 9:55
  SPEAKER_01: Those three mostly fund their data centers from their own cash flow.
• 9:59
  SPEAKER_01: CoreWeave is leveraged on borrowed money.
• 10:01
  SPEAKER_00: Is there any sign demand itself is slowing, not just the money getting shakier?
• 10:05
  SPEAKER_01: Yes.
• 10:06
  SPEAKER_01: Enterprise AI spending growth is already slowing as customers get more efficient and cut costs.
• 10:12
  SPEAKER_01: That's exactly what makes investors nervous about all that leased capacity getting filled.
• 10:16
  SPEAKER_00: I don't think slowing demand is actually the risk here.
• 10:19
  SPEAKER_00: Debt-funded infrastructure bets usually survive because
• 10:23
  SPEAKER_00: the bigger players just absorb the smaller ones.
• 10:25
  SPEAKER_01: Maybe, but that's a bet on someone else bailing out a $72 billion position, not a guarantee.
• 10:31
  SPEAKER_00: Fair.
• 10:32
  SPEAKER_00: It's a bet either way.
• 10:33
  SPEAKER_01: What's new today isn't the narrative,
• 10:35
  SPEAKER_01: it's that credit markets are pricing it in real numbers now,
• 10:38
  SPEAKER_01: not just analyst opinion pieces.
• 10:40
  SPEAKER_00: So what does someone actually do with that if they're not a CoreWeave bondholder?
• 10:44
  SPEAKER_01: If you're choosing or renewing a GPU cloud contract this quarter,
• 10:47
  SPEAKER_01: find out whether your vendor owns its hardware outright or leased
• 10:51
  SPEAKER_01: with borrowed money.
• 10:52
  SPEAKER_01: That debt load is exactly what the bond market is now pricing as risk.
• 10:56
  SPEAKER_00: That's a very different procurement question than the one most teams ask.
• 10:59
  SPEAKER_01: Most teams ask about price per GPU hour.
• 11:02
  SPEAKER_01: Maybe they should also ask who's holding the debt behind it.
• 11:05
  SPEAKER_00: Price per hour was never the whole question.
• 11:09
  SPEAKER_01: Three stories, one thread.
• 11:10
  SPEAKER_01: The ladder for junior engineers is getting shorter,
• 11:13
  SPEAKER_01: the models replacing that work are getting cheaper to run,
• 11:16
  SPEAKER_01: and the money bankrolling all of it is now priced like a risky bet,
• 11:20
  SPEAKER_01: not a sure thing.
• 11:21
  SPEAKER_00: I'll give it until next earnings season.
• 11:23
  SPEAKER_00: CoreWeave either renegotiates those leases or a bigger player buys a stake to backstop it.
• 11:28
  SPEAKER_01: Ask me again in the spring.
• 11:30
  SPEAKER_01: That's Monday.
• 11:31
  [music]