Latest / The AI Practitioner Brief / METR's 2x Speedup and Data Center Backlash — November 6, 2026

Transcript

METR's 2x Speedup and Data Center Backlash — November 6, 2026 · The AI Practitioner Brief

• 0:05
  SPEAKER_00: Welcome to The AI Practitioner Brief, the podcast from AIpractitioner.substack.com,
• 0:10
  SPEAKER_00: where I turn the latest AI and tech news into practical insights
• 0:13
  SPEAKER_00: for AI practitioners.
• 0:15
  SPEAKER_00: I'm Lina Faik, an ML engineer, consultant,
• 0:17
  SPEAKER_00: and lecturer in advanced machine learning at Sorbonne University.
• 0:21
  SPEAKER_00: It's Friday, November 6th.
• 0:23
  SPEAKER_01: Morning, Lina. Good way to end the week.
• 0:25
  SPEAKER_00: Whitney is Jimmy, my co-host on the brief who asks the questions I'd rather skip past.
• 0:29
  SPEAKER_01: I do my best to earn that today.
• 0:32
  SPEAKER_00: Good, because this week voters in 10 states shut down new data centers at the ballot box.
• 0:37
  SPEAKER_00: And in the same week GitHub rebuilt its storage system
• 0:40
  SPEAKER_00: to handle 5,000 AI coding agents hitting one repository at once.
• 0:45
  SPEAKER_00: One side of the AI boom just got told no.
• 0:48
  SPEAKER_00: The other side is still building like nobody said anything.
• 0:51
  SPEAKER_01: That's a strange split. What else happened?
• 0:54
  SPEAKER_00: We'll do the headlines first, then slow down for three stories.
• 0:57
  SPEAKER_00: One puts an actual number on how much AI now speeds up AI research,
• 1:02
  SPEAKER_00: one is that data center backlash in full, and one is what GitHub just rebuilt,
• 1:06
  SPEAKER_00: and why.
• 1:07
  SPEAKER_01: Let's get into it.
• 1:08
  SPEAKER_00: OpenAI released GPT-6 Cyber,
• 1:10
  SPEAKER_00: a model built to find and help patch software vulnerabilities
• 1:13
  SPEAKER_00: while staying inside rules a researcher sets for it.
• 1:17
  SPEAKER_00: In testing, it found 81 new high-severity bugs in Firefox
• 1:21
  SPEAKER_00: and wrote working exploit code for 948 Linux kernel flaws.
• 1:25
  SPEAKER_01: So it's a hacker that OpenAI keeps on a leash.
• 1:28
  SPEAKER_00: Pretty much.
• 1:29
  SPEAKER_00: The company says it stayed inside its assigned scope more than 99% of the time
• 1:34
  SPEAKER_00: without extra safety prompts.
• 1:36
  SPEAKER_00: And with them, it hit 100%.
• 1:38
  SPEAKER_00: Access is rolling out gradually to vetted security customers.
• 1:42
  SPEAKER_01: Announced. Not open to everyone yet.
• 1:44
  SPEAKER_00: Right. Thinking Machines Lab, a startup that builds AI training infrastructure,
• 1:49
  SPEAKER_00: also released Inkling Too Small, a model with 594 billion total parameters,
• 1:54
  SPEAKER_00: but only 17 billion of them active on any given request.
• 1:57
  SPEAKER_01: That gap between total and active, that's the mixture of experts trick, right?
• 2:03
  SPEAKER_00: Right. Mixture of experts means the model splits into specialist subnetworks
• 2:07
  SPEAKER_00: and only wakes up the ones it needs for a given question,
• 2:10
  SPEAKER_00: which is how a huge model still runs cheaply.
• 2:13
  SPEAKER_00: It's available now through Tinker, the company's own training platform,
• 2:16
  SPEAKER_00: with open waits promised later once safety testing wraps up.
• 2:20
  SPEAKER_01: Later meaning no date yet.
• 2:21
  SPEAKER_00: No date yet. That's the AI side on the model front.
• 2:25
  SPEAKER_00: Anthropic's own IPO is in the news too.
• 2:28
  SPEAKER_00: Potential investors are telling reporters they doubt Anthropic,
• 2:31
  SPEAKER_00: the company behind the Claude chatbot,
• 2:33
  SPEAKER_00: can hit the $1.5 trillion valuation it wanted,
• 2:37
  SPEAKER_00: with some now talking about a figure closer to $1 trillion.
• 2:40
  SPEAKER_01: What spooked them?
• 2:41
  SPEAKER_00: The Nasdaq dropped about 6% this week,
• 2:44
  SPEAKER_00: erasing roughly $2.5 trillion in tech stock value right after the midterm elections.
• 2:49
  SPEAKER_00: Anthropic also left its October revenue number blank in its IPO filing,
• 2:54
  SPEAKER_00: which investors are watching closely.
• 2:57
  SPEAKER_01: That's the AI side done.
• 2:59
  SPEAKER_01: What's happening elsewhere in tech?
• 3:01
  SPEAKER_00: Google engineers published a write-up on an outage where the Kubernetes API,
• 3:05
  SPEAKER_00: the control system that manages containerized applications,
• 3:08
  SPEAKER_00: went silent for some users.
• 3:10
  SPEAKER_00: They traced it to three separate failures stacked on top of each other.
• 3:14
  SPEAKER_00: A networking bug, software that wouldn't close its connections properly,
• 3:17
  SPEAKER_00: and a bug in the Linux kernel itself.
• 3:20
  SPEAKER_01: Three bugs at once.
• 3:21
  SPEAKER_01: That's a nightmare to debug.
• 3:22
  SPEAKER_00: It took three layers of tracing to find all three.
• 3:25
  SPEAKER_00: Two infrastructure moves landed the same week.
• 3:28
  SPEAKER_00: Google donated three projects for running AI agents safely inside containers
• 3:32
  SPEAKER_00: to a non-profit foundation that maintains open source cloud tools
• 3:36
  SPEAKER_00: and a company called Spacelift, which sells infrastructure automation software,
• 3:40
  SPEAKER_00: shipped a tool that lets AI agents search OpenTofu's documentation,
• 3:45
  SPEAKER_00: OpenTofu being an open source alternative to the infrastructure tool Terraform,
• 3:49
  SPEAKER_00: without being able to run any commands themselves.
• 3:52
  SPEAKER_01: Agents that can look but not touch.
• 3:54
  SPEAKER_00: That's the idea, yes.
• 3:56
  SPEAKER_00: Let's slow down for the story that tries to put an actual number
• 3:59
  SPEAKER_00: on something everyone's been guessing about for two years,
• 4:02
  SPEAKER_00: how much AI now speeds up AI research itself.
• 4:05
  SPEAKER_01: I've heard founders claim 10 times, 20 times, I want to see where this one actually comes from.
• 4:10
  SPEAKER_00: The number comes from METEAR, which stands for Model Evaluation and Threat Research,
• 4:15
  SPEAKER_00: a non-profit that independently checks
• 4:17
  SPEAKER_00: whether the people building frontier AI systems are still in control of them.
• 4:21
  SPEAKER_00: METEAR spent 11 weeks this fall with direct access inside Anthropic
• 4:25
  SPEAKER_00: and concluded its models are now improving on well-defined tasks at roughly twice
• 4:29
  SPEAKER_00: the rate they did between early 2024 and late 2025.
• 4:34
  SPEAKER_01: Twice as fast as before.
• 4:36
  SPEAKER_01: How did they measure that instead of just taking Anthropic's word for it?
• 4:39
  SPEAKER_00: They tracked a capability benchmark across every model Anthropic released
• 4:43
  SPEAKER_00: since 2024, reviewed internal data on training runs and compute use,
• 4:48
  SPEAKER_00: interviewed about 15 researchers,
• 4:50
  SPEAKER_00: and ran company-wide surveys asking people how much AI tools were speeding up their
• 4:55
  SPEAKER_00: own work.
• 4:56
  SPEAKER_01: And the answer was twice as fast across the board?
• 4:58
  SPEAKER_00: Roughly, with a wide range.
• 5:00
  SPEAKER_00: Mitter's own confidence interval runs from 1.3 times to 3 times,
• 5:04
  SPEAKER_00: because the uncertainty is genuinely large.
• 5:07
  SPEAKER_00: The central estimate is that Anthropik is now getting through about 2 years
• 5:11
  SPEAKER_00: of its old pace of progress in a single year.
• 5:13
  SPEAKER_01: What's actually driving it? More GPUs?
• 5:16
  SPEAKER_01: More people?
• 5:17
  SPEAKER_01: Or the AI itself?
• 5:18
  SPEAKER_00: Not more training compute.
• 5:20
  SPEAKER_00: That scales normally.
• 5:21
  SPEAKER_00: Meteor attributes around 80% of the acceleration
• 5:24
  SPEAKER_00: to researchers actually using AI tools in their daily work,
• 5:28
  SPEAKER_00: based on what Anthropics staff self-reported.
• 5:30
  SPEAKER_01: Self-reported productivity numbers.
• 5:32
  SPEAKER_01: That's the part I don't trust.
• 5:34
  SPEAKER_00: Meteor flags that too.
• 5:36
  SPEAKER_00: They call the survey data loose evidence,
• 5:38
  SPEAKER_00: and they're upfront that one figure going around,
• 5:40
  SPEAKER_00: a claimed 12 times productivity uplift from Anthropik's newest model,
• 5:44
  SPEAKER_00: is exactly that kind of low-effort self-report, not a measured result.
• 5:48
  SPEAKER_01: So the headline multiplier and the number behind it aren't the same level of solid.
• 5:52
  SPEAKER_00: Right, and there's a bigger gap than that.
• 5:54
  SPEAKER_00: Meteor wrote directly that while they could share their top-level conclusions,
• 5:58
  SPEAKER_00: Anthropik redacted most of the evidence Meteor considered most informative
• 6:02
  SPEAKER_00: for those conclusions.
• 6:03
  SPEAKER_01: So the part of the report that would actually let you check their math, we don't get to see?
• 6:07
  SPEAKER_00: Correct.
• 6:08
  SPEAKER_00: Meteor stands by the headline number, but the strongest evidence behind it isn't public.
• 6:13
  SPEAKER_01: Does this apply to any team using AI tools,
• 6:15
  SPEAKER_01: or just a frontier lab with this much internal access?
• 6:18
  SPEAKER_00: Just the one lab, so far.
• 6:20
  SPEAKER_00: Meteor only had this access inside Anthropic, with tooling and feedback loops built over years.
• 6:26
  SPEAKER_00: There's no comparison group,
• 6:27
  SPEAKER_00: so we don't know if a team without that infrastructure sees anything close.
• 6:32
  SPEAKER_01: Then what do you actually do with a number like that?
• 6:35
  SPEAKER_00: It's the first rigorous attempt to measure something every lab has claimed informally.
• 6:39
  SPEAKER_00: Mater found no bottleneck that would stop it compounding,
• 6:42
  SPEAKER_00: which means the gap between labs with this feedback loop and labs
• 6:45
  SPEAKER_00: without it gets wider every quarter.
• 6:48
  SPEAKER_01: And closing that gap needs more than smarter researchers.
• 6:51
  SPEAKER_01: It needs power and chips.
• 6:53
  SPEAKER_00: Which is exactly where this collides with something else that happened this same week.
• 6:56
  SPEAKER_01: What happened?
• 6:57
  SPEAKER_00: On Tuesday, voters in 10 states decided 41 separate ballot measures
• 7:01
  SPEAKER_00: about data centers, the giant server farms that run AI.
• 7:05
  SPEAKER_00: By Wednesday, voters had sided with the industry in only one of them.
• 7:09
  SPEAKER_01: Only one out of 41 passed in their favor?
• 7:12
  SPEAKER_00: That's right.
• 7:13
  SPEAKER_00: In at least 38 of those jurisdictions, voters rejected or restricted new construction.
• 7:19
  SPEAKER_00: Ohio alone had 21 of the 41 measures on its ballots.
• 7:22
  SPEAKER_01: Is this the first time voters have pushed back like this?
• 7:25
  SPEAKER_00: No.
• 7:26
  SPEAKER_00: There were six earlier votes this year, in New Hampshire, Wisconsin, Michigan, and California.
• 7:31
  SPEAKER_00: And every single one also went against new data center construction.
• 7:35
  SPEAKER_00: This week was just the biggest batch at once.
• 7:38
  SPEAKER_01: Any one of those stick out?
• 7:39
  SPEAKER_00: Wells, Maine.
• 7:40
  SPEAKER_00: More than 80% of voters there approved a retroactive 180-day moratorium
• 7:45
  SPEAKER_00: on new data center construction, which pauses even projects already underway.
• 7:49
  SPEAKER_01: 80% isn't a close vote.
• 7:51
  SPEAKER_01: What are people actually angry about?
• 7:54
  SPEAKER_00: Mostly the physical footprint.
• 7:55
  SPEAKER_00: Data centers draw heavily on local power grids and water supplies,
• 7:59
  SPEAKER_00: and construction disrupts small towns fast.
• 8:02
  SPEAKER_01: Why does one data center strain a whole town's grid that much?
• 8:05
  SPEAKER_00: A single large AI data center can draw as much electricity as a small city,
• 8:09
  SPEAKER_00: because training and running large models means running tens of thousands
• 8:13
  SPEAKER_00: of chips continuously, not switching on and off like ordinary computing.
• 8:17
  SPEAKER_01: That explains why a 180-day moratorium is such a big deal for a project already
• 8:22
  SPEAKER_01: under construction.
• 8:23
  SPEAKER_00: Exactly. It can stall equipment deliveries and financing timelines
• 8:26
  SPEAKER_00: that assumed construction wouldn't stop.
• 8:29
  SPEAKER_00: This all ended the same week the Nasdaq dropped 6%
• 8:32
  SPEAKER_00: and wiped out roughly $2.5 trillion in tech stock value,
• 8:36
  SPEAKER_00: so the timing made the backlash louder.
• 8:38
  SPEAKER_01: I think this has less to do with AI specifically and more to do with people's power bills.
• 8:43
  SPEAKER_01: AI's just the thing plugged into the meter right now.
• 8:46
  SPEAKER_00: That's a fair read.
• 8:47
  SPEAKER_00: The industry's response backs it up.
• 8:49
  SPEAKER_00: Companies are offering towns direct payments, school funding,
• 8:53
  SPEAKER_00: and promises to bring their own power generation instead of pulling
• 8:56
  SPEAKER_00: from the local grid.
• 8:57
  SPEAKER_01: Does any of that change the federal picture, or is this purely local?
• 9:01
  SPEAKER_00: Still local for now.
• 9:02
  SPEAKER_00: These are zoning and permitting votes, not federal policy.
• 9:05
  SPEAKER_00: So a company blocked in one county can often still build in the next one over.
• 9:09
  SPEAKER_01: So it doesn't actually fix the compute shortage everyone's been talking about?
• 9:13
  SPEAKER_00: No. If anything, it adds friction right when AI labs need more capacity, not less.
• 9:18
  SPEAKER_00: Which makes it strange that the same week,
• 9:21
  SPEAKER_00: GitHub was busy rebuilding its own infrastructure, as if none of this mattered.
• 9:25
  SPEAKER_01: What did GitHub actually do?
• 9:26
  SPEAKER_00: GitHub replaced Spokes, the storage system it had run for about 15 years
• 9:31
  SPEAKER_00: with a new architecture called Rubicon,
• 9:33
  SPEAKER_00: built specifically to handle AI coding agents pushing changes at once.
• 9:37
  SPEAKER_01: Why did 15-year-old infrastructure suddenly become a problem?
• 9:40
  SPEAKER_00: Traffic.
• 9:41
  SPEAKER_00: GitHub's monthly activity more than doubled between September 2025 and August 2026,
• 9:47
  SPEAKER_00: from about 218 billion events to 473 billion.
• 9:51
  SPEAKER_00: Commits alone hit $7.4 billion in September 2026, five times what they were a year earlier.
• 9:58
  SPEAKER_01: That's the agents, writing code around the clock.
• 10:01
  SPEAKER_00: That's the read, yes.
• 10:02
  SPEAKER_00: The strain showed up as outages ten performance-degrading incidents in April, nine more in May.
• 10:08
  SPEAKER_01: So Spokes just couldn't keep up.
• 10:10
  SPEAKER_01: What was actually wrong with it?
• 10:11
  SPEAKER_00: Spokes stored full copies of a repository across multiple disks and needed a quorum
• 10:16
  SPEAKER_00: of those copies to confirm before a write counted as done.
• 10:19
  SPEAKER_00: Every push was only as fast as the slowest disk in that quorum.
• 10:23
  SPEAKER_01: So one slow disk held up everyone pushing to that repo.
• 10:26
  SPEAKER_00: Right. Rubicon splits the job in two.
• 10:29
  SPEAKER_00: Code changes still go through a fast replicated log for coordination,
• 10:33
  SPEAKER_00: but the file data itself writes once to cloud storage instead of waiting
• 10:37
  SPEAKER_00: on multiple disks to agree.
• 10:38
  SPEAKER_01: Does reading the code change too?
• 10:40
  SPEAKER_00: Yes. Reads now go through separate lightweight workers that cache data,
• 10:44
  SPEAKER_00: so a flood of agents reading a repo doesn't compete with agents writing to it.
• 10:49
  SPEAKER_01: What did that buy them in testing?
• 10:51
  SPEAKER_00: GitHub's own internal tests showed a 35 times improvement in write throughput,
• 10:56
  SPEAKER_00: and the system handled 5,000 agents working on one repository at once,
• 11:00
  SPEAKER_00: pushing 7,200 times a second.
• 11:03
  SPEAKER_01: Is that running for every repo right now or still a lab result?
• 11:06
  SPEAKER_00: Still internal testing.
• 11:08
  SPEAKER_00: GitHub announced the architecture and published the numbers but hasn't given
• 11:11
  SPEAKER_00: a timeline for rolling it out to everyone.
• 11:13
  SPEAKER_01: So nothing changes for most developers yet.
• 11:16
  SPEAKER_00: Not yet.
• 11:17
  SPEAKER_00: But it tells you what GitHub thinks the next few years of software development look like.
• 11:21
  SPEAKER_00: Fewer humans typing, more agents pushing commits nobody watches in real time.
• 11:29
  SPEAKER_01: Which is exactly why the place those commits land needs to be boring and reliable, not clever.
• 11:34
  SPEAKER_00: That's the trade they made.
• 11:36
  SPEAKER_00: Simpler storage, dumber in places, built to survive volume rather than to be elegant.
• 11:41
  SPEAKER_00: That's the shape of the week.
• 11:42
  SPEAKER_00: Anthropic may be accelerating AI research twice as fast as before.
• 11:47
  SPEAKER_00: Voters just told the industry building the power behind that to slow down.
• 11:50
  SPEAKER_00: And GitHub is rebuilding its own plumbing, on the bet that agents keep multiplying regardless.
• 11:55
  SPEAKER_01: A year from now, I don't think the ballot measures change the math.
• 11:59
  SPEAKER_01: Power stays tight, more towns vote exactly like Wells did,
• 12:03
  SPEAKER_01: and the labs just build somewhere quieter.
• 12:05
  SPEAKER_00: If you run agents against GitHub at real volume,
• 12:08
  SPEAKER_00: check now whether your pipeline assumes the old push limits because Rubicon is
• 12:12
  SPEAKER_00: about to move them.
• 12:13
  SPEAKER_00: That's Friday.
• 12:14
  [music]