Latest / The AI Practitioner Brief / Meta's Memory Agent, Anthropic's AI Worm, and Maia 300 — October 13, 2026

Transcript

Meta's Memory Agent, Anthropic's AI Worm, and Maia 300 — October 13, 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,
• 0:10
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:14
  SPEAKER_00: for AI practitioners.
• 0:16
  SPEAKER_00: I'm Lina Faik,
• 0:17
  SPEAKER_00: an ML engineer,
• 0:18
  SPEAKER_00: consultant, and lecturer in advanced machine learning at Sorbonne University.
• 0:23
  SPEAKER_00: It's Tuesday, October 13.
• 0:25
  SPEAKER_01: Tuesday already, and it's a loaded one.
• 0:27
  SPEAKER_00: With me is Jimmy, my co-host on The Brief, who asks the questions I skipped past.
• 0:32
  SPEAKER_00: Today Meta released a model built to remember everything about you and keep working
• 0:36
  SPEAKER_00: on hard problems without giving up.
• 0:39
  SPEAKER_00: And on the very same day,
• 0:40
  SPEAKER_00: Anthropik published research warning that kind
• 0:43
  SPEAKER_00: of persistence could let AI copy itself onto smart ovens and TV boxes
• 0:47
  SPEAKER_00: without anyone's permission.
• 0:49
  SPEAKER_01: So the same week one company sells autonomy as a feature,
• 0:53
  SPEAKER_01: another one is explaining why autonomy is also the risk.
• 0:56
  SPEAKER_00: That's the shape of today.
• 0:58
  SPEAKER_00: We'll start with the headlines,
• 0:59
  SPEAKER_00: then three deep dives,
• 1:01
  SPEAKER_00: Meta just shipped a new agent model,
• 1:03
  SPEAKER_00: Anthropic is warning that AI could copy itself onto smart devices,
• 1:07
  SPEAKER_00: and Microsoft is betting big on its own AI chips.
• 1:10
  SPEAKER_01: Let's get into it.
• 1:12
  SPEAKER_00: First, a new standard called WebMCP is getting attention today.
• 1:16
  SPEAKER_00: MCP, short for Model Context Protocol,
• 1:19
  SPEAKER_00: is the standard way of wiring outside tools into an AI model.
• 1:23
  SPEAKER_00: WebMCP extends that to websites,
• 1:25
  SPEAKER_00: letting a site hand an agent direct software functions instead
• 1:29
  SPEAKER_00: of making it click around the page like a person would.
• 1:32
  SPEAKER_01: So agents stop guessing at buttons on a screen and just call a function directly?
• 1:36
  SPEAKER_00: Exactly,
• 1:37
  SPEAKER_00: and the pitch is that it's faster and cheaper than having
• 1:39
  SPEAKER_00: the agent visually interpret a page every time.
• 1:42
  SPEAKER_01: That's a real cost saving if it catches on.
• 1:45
  SPEAKER_00: Separately,
• 1:45
  SPEAKER_00: some users
• 1:46
  SPEAKER_00: on Anthropix Cloud Opus 5.5 are reporting unusually sharp answers this week.
• 1:51
  SPEAKER_00: The pattern suggests some traffic is quietly being routed to an unreleased,
• 1:55
  SPEAKER_00: more capable checkpoint being tested live.
• 1:58
  SPEAKER_01: A stealth preview nobody signed up for.
• 2:01
  SPEAKER_00: That's the read, though Anthropic hasn't confirmed it.
• 2:03
  SPEAKER_00: That's the AI side.
• 2:05
  SPEAKER_00: Elsewhere in tech today,
• 2:06
  SPEAKER_00: Amazon now runs more than one million warehouse robots and plans to add many more.
• 2:11
  SPEAKER_00: Executives say that eases the physical strain on workers,
• 2:15
  SPEAKER_00: but labor advocates argue people still have to keep pace with the machines.
• 2:19
  SPEAKER_01: Those two claims aren't really in conflict, they're just told from different sides of the floor.
• 2:24
  SPEAKER_00: Epic Games,
• 2:25
  SPEAKER_00: the studio behind Fortnite,
• 2:26
  SPEAKER_00: is reportedly shifting to an AI-first approach for its collaboration with Disney,
• 2:31
  SPEAKER_00: where AI generates code and art that human developers then manage.
• 2:35
  SPEAKER_00: Epic says the goal is more ambitious games, not fewer jobs after layoffs back in March.
• 2:40
  SPEAKER_01: Convenient timing for that explanation.
• 2:42
  SPEAKER_00: And X, the social network formerly called Twitter, is having a moment again.
• 2:47
  SPEAKER_00: Nvidia's Jensen Huang,
• 2:49
  SPEAKER_00: Microsoft's Satya Nadella and Meta's Mark Zuckerberg are all posting there now,
• 2:53
  SPEAKER_00: because it's still the fastest way to get a message out in tech circles.
• 2:57
  SPEAKER_01: Zuckerberg posting there today isn't a coincidence either.
• 3:00
  SPEAKER_01: Meta picked today to drop its biggest AI release in months.
• 3:04
  SPEAKER_00: Right, let's start there.
• 3:05
  SPEAKER_00: Meta released two new models today, News Wonder and Wonderful.
• 3:09
  SPEAKER_00: The company calls them Frontier Class,
• 3:11
  SPEAKER_00: meaning they compete with the best models anyone is currently shipping,
• 3:15
  SPEAKER_00: and they're built for AI agents that work for you personally,
• 3:18
  SPEAKER_00: not just answer questions.
• 3:20
  SPEAKER_01: Frontier class is doing a lot of work in that sentence.
• 3:23
  SPEAKER_01: What can they actually do differently?
• 3:24
  SPEAKER_00: Two things stand out.
• 3:26
  SPEAKER_00: Active memory means the model carries real context about you across conversations
• 3:30
  SPEAKER_00: and actively pulls up what's relevant instead of forgetting after every chat.
• 3:35
  SPEAKER_00: Active focus means it keeps working on a hard goal,
• 3:38
  SPEAKER_00: trying different approaches and asking you for clarification instead
• 3:41
  SPEAKER_00: of just giving up or guessing.
• 3:43
  SPEAKER_01: So it won't quietly hand you a wrong answer and move on.
• 3:46
  SPEAKER_00: That's the pitch.
• 3:47
  SPEAKER_00: Muse Wonder is live now through Meta's API,
• 3:50
  SPEAKER_00: $4 per million input tokens and $20 per million output tokens.
• 3:55
  SPEAKER_00: Wonderful is locked to subscribers of Meta's Muse Code app,
• 3:59
  SPEAKER_00: $50 a month for regular use, $200 for the heaviest tier.
• 4:02
  SPEAKER_01: $200 a month for a chat app is a real number.
• 4:06
  SPEAKER_00: It's a real pivot too.
• 4:07
  SPEAKER_00: Meta's Lama models were open source and mostly chased raw coding and benchmark scores.
• 4:12
  SPEAKER_00: After Lama 4 underwhelmed,
• 4:14
  SPEAKER_00: Meta brought in Alexander Wong,
• 4:16
  SPEAKER_00: the co-founder of Scale AI,
• 4:17
  SPEAKER_00: the data labeling company he built to train other labs models to run
• 4:21
  SPEAKER_00: a new division called Meta Super Intelligence Labs.
• 4:24
  SPEAKER_01: And Wong decided personal assistants beat another coding model?
• 4:27
  SPEAKER_00: That's his bet.
• 4:28
  SPEAKER_00: He said publicly that Meta's real edge isn't winning parameter wars,
• 4:32
  SPEAKER_00: it's personal AI people use every day.
• 4:35
  SPEAKER_00: The first model out of his lab,
• 4:36
  SPEAKER_00: called MuseSpark, was also Meta's first fully closed source model ever.
• 4:41
  SPEAKER_01: Closed?
• 4:42
  SPEAKER_01: That's a big reversal for a company that built its whole AI identity on open weights.
• 4:47
  SPEAKER_00: One actually explained why in Meta's own preparedness report,
• 4:50
  SPEAKER_00: he said early training runs on MuseSpark triggered high-risk flags,
• 4:54
  SPEAKER_00: particularly around bio-risk, and that's part of why they kept it closed.
• 4:58
  SPEAKER_01: So an internal safety test is the reason Meta abandoned open source, not a competitive one.
• 5:03
  SPEAKER_00: That's what the company says, and it's not something anyone outside Meta can check.
• 5:08
  SPEAKER_00: Nobody else has seen what those training runs actually produced.
• 5:11
  SPEAKER_01: I'd want to know what high risk means before taking that at face value.
• 5:15
  SPEAKER_01: That phrase covers a lot of ground.
• 5:17
  SPEAKER_00: Fair.
• 5:18
  SPEAKER_00: Mehta hasn't published the detail,
• 5:20
  SPEAKER_00: so we don't know if that's genuinely dangerous chemistry knowledge
• 5:23
  SPEAKER_00: or something far more ordinary that still tripped a conservative filter.
• 5:26
  SPEAKER_01: Does any of this change what Muse Wonder can actually do for someone building agents today?
• 5:31
  SPEAKER_00: If active memory works as described,
• 5:33
  SPEAKER_00: it's a real answer to the thing that makes most agent demos fall apart,
• 5:37
  SPEAKER_00: which is the model for getting what you told it three steps ago.
• 5:40
  SPEAKER_00: That's the gap between a chatbot and something you trust with a real task.
• 5:44
  SPEAKER_01: How does that stack up against what OpenAI and Anthropic already do with memory?
• 5:49
  SPEAKER_00: Both already carry some memory across sessions.
• 5:51
  SPEAKER_00: What's new here is pairing that with Active Focus so
• 5:54
  SPEAKER_00: the model doesn't just remember,
• 5:56
  SPEAKER_00: it keeps working a hard problem instead of bailing at the first obstacle.
• 5:59
  SPEAKER_01: A model that remembers everything about you and keeps working
• 6:03
  SPEAKER_01: on its own is exactly what makes autonomy feel like a feature.
• 6:06
  SPEAKER_01: It's also exactly the thing you'd want a kill switch for.
• 6:09
  SPEAKER_00: Which is the uncomfortable question underneath today's other big AI story.
• 6:13
  SPEAKER_00: Anthropic just published research asking how close current models already are
• 6:17
  SPEAKER_00: to spreading themselves without anyone's permission at all.
• 6:20
  SPEAKER_01: Close to what, exactly?
• 6:22
  SPEAKER_01: What does self-replicating AI mean in practice?
• 6:25
  SPEAKER_00: Anthropic broke it into four steps.
• 6:27
  SPEAKER_00: Find a vulnerable device,
• 6:29
  SPEAKER_00: get code running on it,
• 6:30
  SPEAKER_00: get the model doing inference there, then repeat the process from that new device.
• 6:34
  SPEAKER_00: It's less an AI takeover and more a worm,
• 6:37
  SPEAKER_00: the old kind of self-spreading malicious program,
• 6:40
  SPEAKER_00: except it's carrying a language model instead of a simple payload.
• 6:43
  SPEAKER_01: What did they actually test it on?
• 6:45
  SPEAKER_00: A model called Gemma 4,
• 6:46
  SPEAKER_00: an open model from Google,
• 6:48
  SPEAKER_00: meaning anyone can download and run it, with 26 billion parameters.
• 6:52
  SPEAKER_00: They ran it across 15 cheap Android TV boxes, about $55 each.
• 6:57
  SPEAKER_00: It only managed 0.39 tokens per second, barely crawling, but it ran.
• 7:01
  SPEAKER_01: That's painfully slow.
• 7:03
  SPEAKER_01: Does slow still count as a real threat?
• 7:05
  SPEAKER_00: Speed wasn't the finding.
• 7:06
  SPEAKER_00: The finding is that a $55 box could host it at all.
• 7:10
  SPEAKER_00: A separate open model,
• 7:11
  SPEAKER_00: GLM-5.2,
• 7:13
  SPEAKER_00: from the Chinese AI lab Zhipu, completed the full four-step chain end-to-end.
• 7:17
  SPEAKER_01: How reliably did it pull that off?
• 7:19
  SPEAKER_00: In Anthropics tests,
• 7:20
  SPEAKER_00: it spread to all eight target devices in four out of five trials,
• 7:24
  SPEAKER_00: and seven of eight in the fifth.
• 7:26
  SPEAKER_00: That's a model running the entire replication chain on its own, repeatedly.
• 7:30
  SPEAKER_01: Was the environment realistic or did they make it easy for the thing to succeed?
• 7:34
  SPEAKER_00: They made the breaking step deliberately trivial.
• 7:36
  SPEAKER_00: The devices had no password on their remote access port,
• 7:39
  SPEAKER_00: basically an unlocked backdoor anthropic left open on purpose,
• 7:43
  SPEAKER_00: so they call this an upper bound, not a real world result.
• 7:46
  SPEAKER_01: So real devices with real passwords might stop this cold.
• 7:50
  SPEAKER_00: Possibly, and that's exactly the uncertainty Anthropic flags.
• 7:53
  SPEAKER_00: We don't know how many consumer devices are actually this exposed,
• 7:57
  SPEAKER_00: only that plenty of cheap internet connected gadgets,
• 8:00
  SPEAKER_00: ovens and TV boxes included, ship with weak defaults.
• 8:03
  SPEAKER_01: Why does Anthropic's own safety team call this hard to stop?
• 8:07
  SPEAKER_00: Because there's no central server to shut down,
• 8:10
  SPEAKER_00: every infected device is its own copy, so taking down one doesn't touch the others.
• 8:15
  SPEAKER_00: Anthropic's Red Team,
• 8:16
  SPEAKER_00: the group that stress tests its own models for danger,
• 8:18
  SPEAKER_00: called it challenging to shut down for exactly that reason.
• 8:21
  SPEAKER_01: That's the same problem security teams already have with botnets,
• 8:25
  SPEAKER_01: networks of hijacked devices used for attacks.
• 8:29
  SPEAKER_01: This just swaps in a model as the payload.
• 8:31
  SPEAKER_00: Right, and that's part of why Anthropic ran this study and published it themselves,
• 8:35
  SPEAKER_00: rather than waiting for someone else to find it first.
• 8:38
  SPEAKER_01: Does this change anything for someone running agents in a normal company, not on a smart oven?
• 8:43
  SPEAKER_00: Not directly, this is about consumer devices, not enterprise agent stacks.
• 8:47
  SPEAKER_00: But the containment question is identical.
• 8:49
  SPEAKER_00: What can your agent actually reach?
• 8:51
  SPEAKER_00: And could it act on that reach without you watching?
• 8:54
  SPEAKER_01: That's worth checking regardless of whether the oven scenario ever happens.
• 8:58
  SPEAKER_00: It's also a reminder that none of this runs for free.
• 9:00
  SPEAKER_00: Every one of these models,
• 9:02
  SPEAKER_00: running or replicating,
• 9:03
  SPEAKER_00: needs power and chips somewhere, which is exactly the bet Microsoft made today.
• 9:07
  SPEAKER_01: Let's hear it.
• 9:08
  SPEAKER_00: Microsoft unveiled its third custom AI chip today, called Maya 300.
• 9:12
  SPEAKER_00: The company says it delivers 40% more tokens per megawatt than the Nvidia GPUs it runs today.
• 9:18
  SPEAKER_01: Tokens per megawatt.
• 9:20
  SPEAKER_01: Translate that for me.
• 9:21
  SPEAKER_00: It's how much AI work a chip squeezes out of a fixed amount of electricity.
• 9:25
  SPEAKER_00: More tokens per megawatt means the same power bill buys more answers.
• 9:29
  SPEAKER_01: And 40% better than Nvidia, according to who?
• 9:32
  SPEAKER_00: According to Microsoft's own testing,
• 9:34
  SPEAKER_00: nobody outside the company has benchmarked a Maya 300 rack yet,
• 9:38
  SPEAKER_00: since it doesn't enter production until 2027.
• 9:41
  SPEAKER_01: How does this compare to Microsoft's last chip?
• 9:43
  SPEAKER_00: Maya 300 packs 96 chips per rack,
• 9:46
  SPEAKER_00: double the number in the previous Maya 200 rack,
• 9:49
  SPEAKER_00: and Microsoft says it pushes through 7 times more tokens per rack overall.
• 9:53
  SPEAKER_00: A full rack draws about 167 kW.
• 9:56
  SPEAKER_01: How does that stack up against what Nvidia itself ships?
• 10:00
  SPEAKER_00: Nvidia's current GB300 racks hold 72 GPUs and draw about 142 kW.
• 10:06
  SPEAKER_00: So Microsoft's rack uses more power total,
• 10:08
  SPEAKER_00: but claims enough extra throughput to still come out a head per watt.
• 10:12
  SPEAKER_01: I don't think this is really about beating NVIDIA on a spec sheet.
• 10:15
  SPEAKER_01: I think it's about Microsoft refusing to let its biggest cost line depend on one supplier.
• 10:21
  SPEAKER_00: That matches how Microsoft is positioning it.
• 10:23
  SPEAKER_00: Maya 300 won't be rented out to cloud customers directly.
• 10:27
  SPEAKER_00: It's for running Microsoft's own products,
• 10:29
  SPEAKER_00: things like its co-pilot assistant and Azure's internal AI workloads.
• 10:33
  SPEAKER_01: So this isn't a product,
• 10:34
  SPEAKER_01: it's Microsoft trying to stop paying NVIDIA's margins on its own infrastructure.
• 10:39
  SPEAKER_00: That's the read, and Microsoft isn't alone.
• 10:41
  SPEAKER_00: Google has built its own TPUs, short for Tensor Processing Units, for years.
• 10:46
  SPEAKER_00: Amazon has its own called Tranium,
• 10:48
  SPEAKER_00: and basically every major cloud company is now building custom chips for
• 10:51
  SPEAKER_00: the same reason, NVIDIA's prices and lead times.
• 10:54
  SPEAKER_01: Does NVIDIA actually lose anything if none of these chips get sold externally?
• 10:59
  SPEAKER_00: Not in the short term,
• 11:00
  SPEAKER_00: Microsoft is still one of Nvidia's biggest customers and will keep buying GPUs for
• 11:04
  SPEAKER_00: the workloads Maya doesn't cover,
• 11:06
  SPEAKER_00: mainly training the largest models rather than running them day to day.
• 11:09
  SPEAKER_01: What don't we actually know here?
• 11:11
  SPEAKER_00: Whether that 40% number holds up once real racks run real customer workloads
• 11:15
  SPEAKER_00: in 2027,
• 11:17
  SPEAKER_00: and how much of Microsoft's total AI capacity actually shifts
• 11:20
  SPEAKER_00: onto its own silicon versus staying on Nvidia.
• 11:23
  SPEAKER_01: Does any of this matter to someone just running inference through Azure right now?
• 11:27
  SPEAKER_00: Not yet directly.
• 11:28
  SPEAKER_00: The chips aren't live until next year,
• 11:30
  SPEAKER_00: but it tells you Microsoft thinks the real constraint on scaling AI is electricity,
• 11:34
  SPEAKER_00: not GPUs, and that's reshaping how fast every cloud provider can grow.
• 11:39
  SPEAKER_01: So that's the day.
• 11:40
  SPEAKER_01: Meta built a model that remembers you and keeps working on hard problems.
• 11:44
  SPEAKER_01: Anthropic is warning that same kind of autonomy could start replicating itself
• 11:48
  SPEAKER_01: on cheap hardware.
• 11:50
  SPEAKER_01: And Microsoft is betting its AI budget on chips built to survive
• 11:54
  SPEAKER_01: the power bill those models run up.
• 11:56
  SPEAKER_00: I don't think the self-replication story is really about ovens.
• 11:59
  SPEAKER_00: I think it's the first real argument for why every company running agents needs
• 12:03
  SPEAKER_00: to audit what those agents can actually reach, not just what they're supposed to do.
• 12:07
  SPEAKER_01: That's the one thing worth doing this week.
• 12:09
  SPEAKER_01: Check what your own agents can actually touch on the network
• 12:13
  SPEAKER_01: before something else finds out first.
• 12:15
  [music]