Latest / The AI Practitioner Brief / Siri's App Access, Anthropic's Proof, and AWS's Speed — November 4, 2026

Transcript

Siri's App Access, Anthropic's Proof, and AWS's Speed — November 4, 2026 · The AI Practitioner Brief

• 0:00
  [music]
• 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 in tech news into practical insights
• 0:14
  SPEAKER_00: for AI practitioners.
• 0:15
  SPEAKER_00: I'm Lina Faik, an ML engineer, consultant,
• 0:18
  SPEAKER_00: and lecturer in advanced machine learning at Sorbonne University.
• 0:21
  SPEAKER_00: It's Wednesday, November 4.
• 0:22
  SPEAKER_01: Wednesday already.
• 0:24
  SPEAKER_01: Feels like Monday was five minutes ago.
• 0:26
  SPEAKER_00: With me is Jimmy, my co-host on The Brief, who asks the questions I'd skip past.
• 0:30
  SPEAKER_01: I'll try to live up to that today.
• 0:31
  SPEAKER_00: Here's the tension today.
• 0:33
  SPEAKER_00: Apple just opened Siri up to every outside app for the first time.
• 0:37
  SPEAKER_00: The same week, OpenAI says ChatGPT alone now handles about a fifth
• 0:41
  SPEAKER_00: of Google's entire search volume.
• 0:43
  SPEAKER_01: So one company's playing catch-up and the other one's already winning without asking permission.
• 0:48
  SPEAKER_00: We'll start with the headlines, then three deep dives.
• 0:50
  SPEAKER_00: Apple is letting every app plug into Siri,
• 0:53
  SPEAKER_00: Anthropic built a way to cryptographically prove which model actually answered you,
• 0:57
  SPEAKER_00: and Amazon made its cloud containers scale four times faster.
• 1:00
  SPEAKER_01: Let's get into it.
• 1:01
  SPEAKER_00: First up, OpenAI.
• 1:03
  SPEAKER_00: The company says ChatGPT's search-like queries have reached about 20% of Google's volume.
• 1:08
  SPEAKER_01: 20% of Google. That's not nothing.
• 1:10
  SPEAKER_00: OpenAI's consumer lead, Adam Fry, said it on November 3rd.
• 1:14
  SPEAKER_00: ChatGPT now has 1.2 billion weekly users, up from 1 billion just two months ago.
• 1:19
  SPEAKER_01: Is that dictation or people actually talking to it?
• 1:21
  SPEAKER_00: Both combined, 250 million weekly voice users.
• 1:25
  SPEAKER_00: OpenAI also says it's planning to merge its chat and work modes into one product.
• 1:29
  SPEAKER_01: Planning. Not done yet.
• 1:30
  SPEAKER_00: Right. That part is still just an announcement.
• 1:33
  SPEAKER_00: Next, Google's Gemini team, Gemini is Google's family of AI models,
• 1:37
  SPEAKER_00: opened a public preview of a notebook API.
• 1:39
  SPEAKER_01: What does a notebook API actually do?
• 1:41
  SPEAKER_00: It lets developers build apps that automatically turn web pages, YouTube videos,
• 1:45
  SPEAKER_00: or plain text into organized notebooks
• 1:47
  SPEAKER_00: the same way Google's own Notebook LM tool does by hand.
• 1:51
  SPEAKER_01: Who gets to use it?
• 1:52
  SPEAKER_00: Right now, only personal Google accounts, not business ones.
• 1:54
  SPEAKER_00: That's the AI side.
• 1:56
  SPEAKER_00: Elsewhere in tech today, Databricks, the data and AI platform company,
• 2:00
  SPEAKER_00: took its serverless Postgres database, called Lakebase,
• 2:03
  SPEAKER_00: to general availability on Google Cloud.
• 2:05
  SPEAKER_01: Remind me what serverless means here.
• 2:07
  SPEAKER_00: You don't provision servers yourself.
• 2:09
  SPEAKER_00: It scales up and down automatically, including down to zero when nothing's running.
• 2:13
  SPEAKER_00: So you're not paying for idle capacity.
• 2:15
  SPEAKER_01: Fine. That's the usual pitch for serverless.
• 2:18
  SPEAKER_00: There's also a new open source tool called Agent Desktop,
• 2:20
  SPEAKER_00: which lets an AI agent click around real desktop apps on Mac, Windows,
• 2:24
  SPEAKER_00: and Linux by reading the app's underlying structure instead of guessing
• 2:28
  SPEAKER_00: from screenshots.
• 2:28
  SPEAKER_01: Honestly, I'm not sure this one earns a headline.
• 2:31
  SPEAKER_01: Feels like a new agent tool like this drops every week.
• 2:34
  SPEAKER_00: Fair, it's a small project, but it's a real alternative to screenshot-based agents,
• 2:38
  SPEAKER_00: worth one line. Last one, a review of recent SaaS filings found that 29 out
• 2:43
  SPEAKER_00: of 48 subscription businesses over $100 million in revenue actually shrank
• 2:47
  SPEAKER_00: in October.
• 2:48
  SPEAKER_01: What was that number in September?
• 2:49
  SPEAKER_00: Four, the analysis points to weaker new sales, not more cancellations,
• 2:54
  SPEAKER_00: and companies under $5 million in revenue were hit hardest.
• 2:56
  SPEAKER_01: Let's start with Siri then, since you opened with it.
• 2:59
  SPEAKER_00: Apple put out a developer note on November 3rd.
• 3:01
  SPEAKER_00: Starting now, in the iOS 27.3 beta, any developer can plug their app into Siri AI.
• 3:07
  SPEAKER_01: Beta meaning not everyone has it yet.
• 3:10
  SPEAKER_00: Right, it's a developer preview, live on the beta versions of iOS, iPadOS, macOS, and VisionOS.
• 3:16
  SPEAKER_00: Regular App Store apps using it won't be allowed until next spring.
• 3:20
  SPEAKER_01: So nobody's actually using this today.
• 3:22
  SPEAKER_00: Developers are testing it today.
• 3:24
  SPEAKER_00: Regular users get it once Apple approves the first wave of apps.
• 3:27
  SPEAKER_00: The mechanism is called AppIntense,
• 3:29
  SPEAKER_00: a framework that lets an app describe what it can do and what content it holds
• 3:33
  SPEAKER_00: so Siri can act on it.
• 3:34
  SPEAKER_01: How is that different from what Siri already does with apps like Maps or Reminders?
• 3:39
  SPEAKER_00: Before, Apple had to handwrite the integration for each of its own apps.
• 3:42
  SPEAKER_00: AppIntense lets any outside developer do that themselves and bundle several steps
• 3:46
  SPEAKER_00: into one request Apple calls a skill.
• 3:49
  SPEAKER_01: Give me an example.
• 3:50
  SPEAKER_00: Say you want to reorder your usual coffee and get it added to your calendar.
• 3:54
  SPEAKER_00: Instead of opening two apps, you ask Siri once, and it calls both apps intense in sequence.
• 3:59
  SPEAKER_01: That's the thing Apple's rivals have had for years, isn't it?
• 4:02
  SPEAKER_00: Pretty much.
• 4:03
  SPEAKER_00: Google and Amazon opened their assistants to outside developers years ago,
• 4:06
  SPEAKER_00: through Google Actions and Alexa skills.
• 4:09
  SPEAKER_00: Apple kept Siri mostly closed,
• 4:11
  SPEAKER_00: partly because it insisted on doing a lot of the processing on the device itself
• 4:14
  SPEAKER_00: for privacy, which is slower to build than running everything in the cloud.
• 4:18
  SPEAKER_01: So this is Apple catching up, not leading.
• 4:20
  SPEAKER_00: That's a fair reading.
• 4:21
  SPEAKER_00: Apple's own personalized Siri upgrade slipped repeatedly
• 4:24
  SPEAKER_00: after first being promised back in 2025.
• 4:27
  SPEAKER_00: Opening Siri to every app is a different move.
• 4:29
  SPEAKER_00: It's admitting Apple can't build all the useful integrations itself.
• 4:33
  SPEAKER_01: Is there a reason it's blocked in the EU?
• 4:35
  SPEAKER_00: The reporting doesn't say why, it's just listed as currently unavailable there,
• 4:39
  SPEAKER_00: along with anyone under 13.
• 4:41
  SPEAKER_00: My guess is regulatory review, but that's a guess, not something Apple has confirmed.
• 4:45
  SPEAKER_01: I'd rather you say that than make something up.
• 4:47
  SPEAKER_00: Exactly, so we're leaving it there.
• 4:49
  SPEAKER_00: What's clear is this is a land grab for attention on the lock screen,
• 4:52
  SPEAKER_00: right as OpenAI's voice mode pulls in a quarter of a billion weekly users
• 4:56
  SPEAKER_00: on its own app.
• 4:57
  SPEAKER_01: So Apple needs other people's apps to make Siri worth opening in the first place.
• 5:01
  SPEAKER_00: That's the bet.
• 5:02
  SPEAKER_00: Whether developers actually billed for it depends on how many
• 5:05
  SPEAKER_00: of these skills Apple approves once spring comes.
• 5:08
  SPEAKER_01: And once all these outside apps are feeding Siri information and taking actions
• 5:11
  SPEAKER_01: through it, that's a bigger permissions question than Apple's note let on.
• 5:15
  SPEAKER_00: It is. And that question, who's allowed to act on your behalf,
• 5:19
  SPEAKER_00: and whether you can trust what actually ran,
• 5:21
  SPEAKER_00: is exactly where the next story picks up.
• 5:23
  SPEAKER_01: Who's allowed to act on your behalf is one trust problem.
• 5:26
  SPEAKER_01: Knowing what actually answered you is another.
• 5:28
  SPEAKER_00: Which is exactly what Anthropic just tackled.
• 5:31
  SPEAKER_00: On November 2nd, Anthropic published FlashProof,
• 5:34
  SPEAKER_00: an open source tool that cryptographically proves which specific AI model generated
• 5:38
  SPEAKER_00: a given response.
• 5:39
  SPEAKER_01: Proves it how?
• 5:40
  SPEAKER_00: It uses something called a ZSK snark,
• 5:42
  SPEAKER_00: a type of cryptographic proof that lets you verify a computation was done correctly
• 5:47
  SPEAKER_00: without redoing it or seeing the private data behind it.
• 5:49
  SPEAKER_01: That sounds like it belongs in cryptocurrency, not AI.
• 5:53
  SPEAKER_00: It does come from the same family of math used in some blockchain systems.
• 5:57
  SPEAKER_00: Here, the model provider runs the request and hands back the answer plus
• 6:00
  SPEAKER_00: a proof tied to a specific, committed version of the model's weights.
• 6:03
  SPEAKER_01: So the proof says this exact model produced this exact answer, not just some model somewhere.
• 6:10
  SPEAKER_00: Right, that's the whole point.
• 6:11
  SPEAKER_00: It's a response to a real complaint that API providers could quietly swap
• 6:15
  SPEAKER_00: in a cheaper, smaller version of a model while still charging for the expensive one,
• 6:19
  SPEAKER_00: and customers have no way to check.
• 6:21
  SPEAKER_01: Has that actually happened, or is this a hypothetical?
• 6:24
  SPEAKER_00: There have been accusations of exactly that across the industry,
• 6:27
  SPEAKER_00: though Anthropik's own announcement doesn't pin it on a specific company.
• 6:30
  SPEAKER_00: The tool is aimed at making that kind of swap checkable.
• 6:33
  SPEAKER_01: Does it actually work yet?
• 6:34
  SPEAKER_00: Anthropic tested it on Meta's open model, Llama 4 Maverick,
• 6:38
  SPEAKER_00: running a shrunk-down 4-bit version with about 400 billion parameters.
• 6:42
  SPEAKER_00: A proof for a thousand token request came out to about 19 megabytes,
• 6:46
  SPEAKER_00: and a normal computer can check it in about 20 seconds.
• 6:48
  SPEAKER_01: 20 seconds to check one answer.
• 6:51
  SPEAKER_01: That's not nothing, but it's not instant either.
• 6:53
  SPEAKER_00: It's not, and that's the catch.
• 6:55
  SPEAKER_00: Generating these proofs
• 6:56
  SPEAKER_00: for a full-sized frontier model costs far more computing power than running
• 7:00
  SPEAKER_00: the model itself.
• 7:01
  SPEAKER_00: Anthropic's own researchers say production use at that scale isn't close
• 7:05
  SPEAKER_00: to commercially viable yet.
• 7:06
  SPEAKER_01: So this doesn't change anything for someone calling an API today?
• 7:10
  SPEAKER_00: Not yet. If you're buying inference right now,
• 7:12
  SPEAKER_00: you still have no receipt proving which model answered you.
• 7:15
  SPEAKER_00: This is a research prototype that shows the idea works, not something live in any product.
• 7:19
  SPEAKER_01: I don't think cost is the only reason this stalls.
• 7:22
  SPEAKER_01: I think a lot of providers would rather you couldn't check in the first place.
• 7:25
  SPEAKER_00: That's your read, not something Anthropics Post says.
• 7:28
  SPEAKER_00: But it's a real incentive problem either way,
• 7:31
  SPEAKER_00: since the providers who'd benefit most from staying unverifiable are
• 7:34
  SPEAKER_00: the least likely to adopt this voluntarily.
• 7:36
  SPEAKER_01: It's open source, though.
• 7:38
  SPEAKER_01: Could someone else build the checking layer and point it at a provider
• 7:41
  SPEAKER_01: that doesn't want to be checked?
• 7:42
  SPEAKER_00: In theory, yes.
• 7:43
  SPEAKER_00: That's exactly why Anthropic released it openly rather than keeping it internal.
• 7:48
  SPEAKER_00: Whether anyone actually does that against an uncooperative provider is unresolved today.
• 7:52
  SPEAKER_01: Feels like the kind of thing that only gets built once someone gets caught.
• 7:55
  SPEAKER_00: Maybe. And that need for proof gets more urgent as more systems,
• 7:59
  SPEAKER_00: like the Siri integrations we just covered,
• 8:01
  SPEAKER_00: start delegating real actions to models nobody outside the company can fully verify.
• 8:06
  SPEAKER_01: Delegating real actions to systems you can't verify?
• 8:09
  SPEAKER_01: Sounds like it needs infrastructure that doesn't fall over, too.
• 8:12
  SPEAKER_00: Which brings us to Amazon.
• 8:13
  SPEAKER_00: AWS's Container Service, called ECS, short for Elastic Container Service,
• 8:18
  SPEAKER_00: got a lot faster at scaling itself up under load.
• 8:21
  SPEAKER_01: Scaling meaning adding more capacity automatically when traffic spikes?
• 8:25
  SPEAKER_00: Exactly. ECS runs your app in containers,
• 8:28
  SPEAKER_00: and Fargate is AWS's option for running those containers without you managing
• 8:32
  SPEAKER_00: the underlying servers.
• 8:33
  SPEAKER_00: The improvement here is about how fast the system notices it needs more capacity.
• 8:37
  SPEAKER_01: What was slow about it before?
• 8:39
  SPEAKER_00: By default, AWS's monitoring service, CloudWatch,
• 8:42
  SPEAKER_00: only publish CPU and memory numbers for your containers once every 60 seconds.
• 8:47
  SPEAKER_00: The autoscaler can only react to data it's actually seen.
• 8:50
  SPEAKER_01: So a traffic spike could sit unnoticed for almost a full minute.
• 8:53
  SPEAKER_00: Right.
• 8:54
  SPEAKER_00: AWS now offers an option to publish those same numbers every 20 seconds instead,
• 8:59
  SPEAKER_00: letting the autoscaler check three times as often.
• 9:01
  SPEAKER_01: And that actually moves the needle.
• 9:03
  SPEAKER_00: According to AWS's own tests,
• 9:05
  SPEAKER_00: the time to even trigger a scale-out dropped from 363 seconds to 86 seconds.
• 9:10
  SPEAKER_00: That's a 76% cut.
• 9:12
  SPEAKER_01: What about actually getting new capacity running, not just noticing you need it?
• 9:16
  SPEAKER_00: Total time from detecting the spike to having new tasks leave dropped from 386 seconds to 109.
• 9:22
  SPEAKER_00: One blog covering it separately found Fargate SLA Breach Time down 69% in its own ramp tests.
• 9:28
  SPEAKER_01: Two different sets of numbers from two different tests.
• 9:31
  SPEAKER_01: Worth keeping those separate.
• 9:32
  SPEAKER_00: Agreed. AWS's benchmark and that blog's benchmark aren't the same test,
• 9:36
  SPEAKER_00: but they point the same direction.
• 9:38
  SPEAKER_01: Is this just free then?
• 9:39
  SPEAKER_01: Seems like an easy switch.
• 9:40
  SPEAKER_00: Not entirely.
• 9:42
  SPEAKER_00: The high-resolution metrics are a new pricing line on your CloudWatch bill,
• 9:45
  SPEAKER_00: so checking three times as often costs more than checking once a minute.
• 9:49
  SPEAKER_01: So you're trading a monitoring bill for fewer missed scaling windows.
• 9:52
  SPEAKER_00: That's the trade exactly.
• 9:54
  SPEAKER_00: For most workloads the extra CloudWatch cost is small next to what
• 9:57
  SPEAKER_00: a missed scaling event costs you in an outage, but it's not nothing.
• 10:00
  SPEAKER_01: I've been on the other end of that 60-second lag before,
• 10:03
  SPEAKER_01: watching a dashboard while a service falls
• 10:05
  SPEAKER_01: over and the autoscaler just hasn't caught up.
• 10:07
  SPEAKER_00: That's the exact gap this closes.
• 10:09
  SPEAKER_00: If your service sees bursty traffic and you're still
• 10:12
  SPEAKER_00: on ECS's default 60-second metrics, this is worth checking this week.
• 10:16
  SPEAKER_00: Switch the target tracking policy to the high-resolution option and weigh
• 10:19
  SPEAKER_00: the CloudWatch cost against how much a slow scale-out actually costs you.
• 10:23
  SPEAKER_01: That's a concrete one.
• 10:24
  SPEAKER_01: Does it apply to everything on ECS?
• 10:26
  SPEAKER_00: It's available across Fargate,
• 10:28
  SPEAKER_00: ECS's managed instances and regular EC2 in every AWS commercial and GovCloud region,
• 10:34
  SPEAKER_00: so there's no setup you're locked out of.
• 10:35
  SPEAKER_01: Good, that's one thing someone can actually go do today.
• 10:38
  SPEAKER_00: Apple opened Siri to outside developers today,
• 10:41
  SPEAKER_00: trying to catch up to a chat GPT that's already winning on its own terms.
• 10:45
  SPEAKER_00: Anthropic gave the industry its first real way to prove which model answered a question.
• 10:49
  SPEAKER_00: And Amazon showed that checking your infrastructure more often pays for itself
• 10:53
  SPEAKER_00: in outages you avoid.
• 10:54
  SPEAKER_01: I'll bet that once Apple's Siri apps actually ship next spring,
• 10:57
  SPEAKER_01: the real story is how many get rejected, not how many get approved.
• 11:01
  SPEAKER_00: That's Wednesday.
• 11:02
  [music]