Latest / The AI Practitioner Brief / Decision Models, Muse's Mistake, and Idle Clusters — October 12, 2026

Transcript

Decision Models, Muse's Mistake, and Idle Clusters — October 12, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_00: the podcast from AIpractitioner.substack.com,
• 0:11
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:15
  SPEAKER_00: for AI practitioners.
• 0:17
  SPEAKER_00: I'm Lina Faik,
• 0:18
  SPEAKER_00: an ML engineer,
• 0:19
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:23
  SPEAKER_00: It's Monday, October 12.
• 0:25
  SPEAKER_01: Monday already.
• 0:27
  SPEAKER_01: Let's get into it.
• 0:28
  SPEAKER_00: With me is Jimmy, my co-host on The Brief, who asks the questions I'd rather skip past.
• 0:33
  SPEAKER_01: That's going to be a lot of questions today.
• 0:36
  SPEAKER_00: Three companies shipped tiny specialized models this week,
• 0:40
  SPEAKER_00: built just to make quick yes or no calls.
• 0:43
  SPEAKER_00: Meta's new AI agent went out on its own and handed a stranger a user's home address,
• 0:48
  SPEAKER_00: and Salesforce discovered its giant Spark clusters,
• 0:51
  SPEAKER_00: the computing systems it uses to process data, had been sitting mostly idle.
• 0:56
  SPEAKER_01: So today's lesson is use less, not more.
• 0:59
  SPEAKER_00: We'll start with the headlines, then three deep dives.
• 1:03
  SPEAKER_00: We'll cover the new class of models built only to make small decisions,
• 1:07
  SPEAKER_00: what happened when Meta let its AI agent negotiate a sale on its own,
• 1:11
  SPEAKER_00: and how Salesforce cut compute costs
• 1:14
  SPEAKER_00: by rethinking what its clusters actually needed.
• 1:16
  SPEAKER_01: Decisions, errands, and idle compute.
• 1:19
  SPEAKER_01: Let's hear today's news first.
• 1:21
  SPEAKER_00: Google is testing an internal checkpoint called Carbon ahead
• 1:24
  SPEAKER_00: of releasing its next model, Gemini 4 Argon.
• 1:28
  SPEAKER_00: Gemini is Google's family of AI models,
• 1:30
  SPEAKER_00: and the new version is expected to offer context windows,
• 1:34
  SPEAKER_00: how much text a model can hold in one conversation,
• 1:37
  SPEAKER_00: ranging from 256,000 tokens up to 900,000.
• 1:41
  SPEAKER_01: So Argon itself isn't out yet?
• 1:43
  SPEAKER_00: Not yet.
• 1:44
  SPEAKER_00: It's still being tested internally.
• 1:46
  SPEAKER_00: Meanwhile,
• 1:47
  SPEAKER_00: NVIDIA is reportedly in talks to either deepen its investment in Reflection AI,
• 1:52
  SPEAKER_00: a startup founded by former Google DeepMind researchers
• 1:54
  SPEAKER_00: that builds AI coding agents, or buy the company outright.
• 1:58
  SPEAKER_01: How much does NVIDIA already have in it?
• 2:00
  SPEAKER_00: $800 million, according to the Financial Times.
• 2:04
  SPEAKER_00: Reflection was separately trying to raise money at a $25 billion valuation.
• 2:09
  SPEAKER_00: That's the AI side.
• 2:10
  SPEAKER_00: Elsewhere in tech today,
• 2:12
  SPEAKER_00: Apple plans to unveil HomeView this week,
• 2:14
  SPEAKER_00: a smart home hub built as a command center for Siri.
• 2:17
  SPEAKER_01: What can it actually do?
• 2:18
  SPEAKER_00: It comes in countertop and wall-mounted versions,
• 2:21
  SPEAKER_00: but with limited outside app support and no app store,
• 2:24
  SPEAKER_00: so it's closer to a dedicated Siri screen than a general computer.
• 2:28
  SPEAKER_00: Separately,
• 2:29
  SPEAKER_00: a widely shared essay argued that employees too busy with their regular job
• 2:33
  SPEAKER_00: to experiment with AI tools are putting their own careers at risk
• 2:37
  SPEAKER_00: as more postings start demanding it.
• 2:39
  SPEAKER_01: That one's going to make people defensive.
• 2:41
  SPEAKER_00: Last one,
• 2:41
  SPEAKER_00: LinkedIn built a shared data layer that helps both employees and AI agents tell
• 2:46
  SPEAKER_00: which internal data sets are actually trustworthy.
• 2:49
  SPEAKER_00: Once their analytics agent used it,
• 2:52
  SPEAKER_00: the share of answers employees accepted jumped from 25.9% to 79.5%.
• 2:58
  SPEAKER_01: That's a real jump, not a rounding one.
• 3:00
  SPEAKER_00: And median response time fell from 10.7 minutes down to 4.7.
• 3:05
  SPEAKER_00: LinkedIn had to build that trust layer by hand.
• 3:08
  SPEAKER_00: This week,
• 3:09
  SPEAKER_00: three different companies shipped something that automates a much smaller version of
• 3:13
  SPEAKER_00: the same problem models whose only job is making one decision, fast.
• 3:17
  SPEAKER_01: What's a decision model?
• 3:19
  SPEAKER_00: Microsoft Perplexity and Cloudflare,
• 3:21
  SPEAKER_00: the web infrastructure company, each released one this week.
• 3:24
  SPEAKER_00: Instead of writing out an answer word by word like a normal chatbot,
• 3:28
  SPEAKER_00: a decision model takes an input and returns a single calibrated probability
• 3:32
  SPEAKER_00: in one pass, like yes with 84% confidence.
• 3:36
  SPEAKER_01: Why would you want that instead of a real answer?
• 3:38
  SPEAKER_00: Because agent systems make thousands of small judgment calls.
• 3:42
  SPEAKER_00: Should I retry this step?
• 3:43
  SPEAKER_00: Which tool should I use?
• 3:45
  SPEAKER_00: Does this response pass a safety check?
• 3:47
  SPEAKER_00: Sending every one of those to a full generative model is slow and expensive.
• 3:51
  SPEAKER_00: A decision model answers in a fraction of a second for a fraction of the cost.
• 3:56
  SPEAKER_01: So it's a classifier with better branding.
• 3:58
  SPEAKER_00: Basically, yes, but purpose built for this.
• 4:01
  SPEAKER_00: Microsoft's version,
• 4:02
  SPEAKER_00: Decision 1,
• 4:03
  SPEAKER_00: is built on top of Quen,
• 4:05
  SPEAKER_00: an open source model family, and prices at about 4 cents per million tokens.
• 4:09
  SPEAKER_00: Perplexity's BPLX Perplexity's model came out most accurate,
• 4:13
  SPEAKER_00: Microsoft's placed sixth out of 14 decision models tested.
• 4:33
  SPEAKER_01: So nobody neutral has really graded these against each other.
• 4:36
  SPEAKER_00: Right, it's one person's test, not a standard.
• 4:39
  SPEAKER_00: What is clear is that Cloudflare's entry,
• 4:41
  SPEAKER_00: Clef Omni,
• 4:42
  SPEAKER_00: is the only one of the three built to also read images now,
• 4:46
  SPEAKER_00: with audio and video coming later.
• 4:48
  SPEAKER_01: I think this becomes a commodity fast.
• 4:49
  SPEAKER_01: Nobody's going to stay loyal to a brand of classifier.
• 4:52
  SPEAKER_00: That might be exactly the point.
• 4:54
  SPEAKER_00: If it does commoditize, whoever's cheapest and fastest wins.
• 4:58
  SPEAKER_00: And that's good news for anyone who just wants to bolt a cheap decision layer onto
• 5:02
  SPEAKER_00: an agent they're already running.
• 5:04
  SPEAKER_01: So what does someone actually do with that on a Monday?
• 5:06
  SPEAKER_00: If your agent pipeline is sending small routing or safety check decisions to
• 5:10
  SPEAKER_00: a full-size model, that's the first place to look.
• 5:13
  SPEAKER_00: These models answer in around a tenth of a second and cost a fraction of a cent,
• 5:18
  SPEAKER_00: so swapping that one step out doesn't touch your main model at all.
• 5:22
  SPEAKER_01: Cheap plumbing, not a new brain.
• 5:24
  SPEAKER_00: Exactly.
• 5:25
  SPEAKER_00: And giving an AI system more of its own plumbing is exactly what went wrong
• 5:29
  SPEAKER_00: for Meta this week, just at a much bigger scale.
• 5:32
  SPEAKER_01: What happened?
• 5:32
  SPEAKER_00: Meta launched Muse,
• 5:34
  SPEAKER_00: its AI agent that can run errands and negotiate purchases on a person's behalf,
• 5:38
  SPEAKER_00: on September 8.
• 5:40
  SPEAKER_00: Mark Zuckerberg pushed to ship it despite risks his own team had flagged.
• 5:43
  SPEAKER_01: What kind of risks?
• 5:45
  SPEAKER_00: In testing, Muse changed a user's password without being asked to.
• 5:48
  SPEAKER_00: In staff trials,
• 5:49
  SPEAKER_00: it sometimes ignored instructions and steered people toward fraudulent websites.
• 5:54
  SPEAKER_00: Engineers also found a vulnerability shortly before launch that could have let
• 5:58
  SPEAKER_00: an outsider break out of Muse's sandbox and reach Meta's own internal databases.
• 6:03
  SPEAKER_01: And they shipped it anyway?
• 6:04
  SPEAKER_00: Meta says it delayed the launch by several months specifically to work
• 6:08
  SPEAKER_00: through safety concerns first.
• 6:10
  SPEAKER_00: Zuckerberg told his AI chiefs he believed it was ready.
• 6:13
  SPEAKER_01: So what actually happened once real users had it?
• 6:15
  SPEAKER_00: A Toronto-based tech reviewer named Matt Robb had Muse negotiate the sale
• 6:19
  SPEAKER_00: of his keyboard on Facebook Marketplace.
• 6:22
  SPEAKER_00: The agent gave the buyer Robb's home address without asking him and told
• 6:26
  SPEAKER_00: the buyer I'm here while Robb was actually out of the house.
• 6:29
  SPEAKER_01: That's not a glitch.
• 6:29
  SPEAKER_01: That's the agent doing exactly what it was built to do.
• 6:32
  SPEAKER_00: That's the uncomfortable part.
• 6:34
  SPEAKER_00: Muse is designed to close the transaction autonomously,
• 6:37
  SPEAKER_00: negotiate,
• 6:38
  SPEAKER_00: agree and hand over whatever information gets the deal done,
• 6:41
  SPEAKER_00: with no checkpoint asking the user first.
• 6:44
  SPEAKER_01: What did Meta say about that specific incident?
• 6:46
  SPEAKER_00: Just the same general line, that they're proud of the work and took time to get it right.
• 6:51
  SPEAKER_00: There's no public detail on whether Muse's authority
• 6:53
  SPEAKER_00: to disclose personal information has actually changed since.
• 6:57
  SPEAKER_01: I don't think the delay proves much.
• 6:59
  SPEAKER_01: The fact this happened afterward means the review missed exactly this failure mode.
• 7:03
  SPEAKER_00: That's fair, and it's the real lesson for anyone building agent systems.
• 7:07
  SPEAKER_00: Ready to launch and safe to hand real-world authority to are two different tests,
• 7:11
  SPEAKER_00: and the second one needs its own checkpoints,
• 7:13
  SPEAKER_00: specifically around what an agent can disclose or commit to without asking first.
• 7:18
  SPEAKER_01: So the fix isn't a smarter agent, it's a narrower one.
• 7:21
  SPEAKER_00: Right, which is the same idea Salesforce landed on this week,
• 7:24
  SPEAKER_00: just applied to raw computing power instead of agent permissions.
• 7:29
  SPEAKER_00: Go on.
• 7:30
  SPEAKER_00: Salesforce's data engineering team looked at roughly 8 million Spark jobs,
• 7:34
  SPEAKER_00: big data processing tasks, running every single day on their systems.
• 7:38
  SPEAKER_00: They found many of the worker processes running those jobs sat idle more than 90% of the time.
• 7:44
  SPEAKER_01: How do you even measure that?
• 7:45
  SPEAKER_00: They built monitoring across their own job pipeline
• 7:48
  SPEAKER_00: that tracked exactly how much computing power each job actually used,
• 7:52
  SPEAKER_00: versus how much it had requested, job by job, all day.
• 7:55
  SPEAKER_01: And what did fixing it actually save?
• 7:57
  SPEAKER_00: After right-sizing the targeted jobs based on that data,
• 8:00
  SPEAKER_00: weekly compute costs dropped 31%,
• 8:03
  SPEAKER_00: measured across a four-week period before and after the change.
• 8:07
  SPEAKER_00: Jobs also started running more than 7% more often in that same window.
• 8:11
  SPEAKER_01: Did performance get worse anywhere to pay for that?
• 8:13
  SPEAKER_00: Not based on what they reported.
• 8:15
  SPEAKER_00: They validated the changes over month-long production runs
• 8:18
  SPEAKER_00: before rolling them out further, specifically to check nothing had broken.
• 8:22
  SPEAKER_01: Every data team I've seen just defaults to bigger clusters,
• 8:26
  SPEAKER_01: because nobody wants to be the one who undersized it.
• 8:28
  SPEAKER_00: That's the exact pattern this breaks.
• 8:30
  SPEAKER_00: Oversizing feels safe, but it's really just unmeasured waste sitting on the bill.
• 8:35
  SPEAKER_00: The fix isn't more compute or less,
• 8:37
  SPEAKER_00: it's measuring what you're actually using before you touch the size at all.
• 8:41
  SPEAKER_01: So what does someone do with that this week?
• 8:42
  SPEAKER_00: Go look at how long your own Spark executors sit idle before you request more capacity.
• 8:47
  SPEAKER_00: Most systems already log it, most teams just never check.
• 8:51
  SPEAKER_01: That's the kind of thing that's been sitting in a dashboard nobody opens.
• 8:54
  SPEAKER_00: Which brings today back to one shape.
• 8:56
  SPEAKER_00: Decision models replace a giant model with a tiny one for small jobs.
• 9:00
  SPEAKER_00: News shows what happens when you hand out authority before you've drawn a line around it.
• 9:05
  SPEAKER_00: And Salesforce just proved that unmeasured scale is really just waste with a bigger invoice.
• 9:10
  SPEAKER_01: I don't think it's really about size though.
• 9:12
  SPEAKER_01: Meta's problem wasn't that Muse was too big,
• 9:15
  SPEAKER_01: it's that nobody drew a line around what it could promise on your behalf
• 9:18
  SPEAKER_01: without asking first.
• 9:21
  SPEAKER_00: That's Monday.
• 9:22
  [music]