AI Digest, Nov 2: Agent Costs, KV Cache, Agent Skills, Tao's Theorem Economy

0:02PM Lead ProThis is PM Lead Pro. Today's digest asks what computer-using AI agents really cost, from GPT-6.1 Terra pricing to KV cache and Agent Skills, plus Terry Tao on AI in mathematics.

0:16AI is a, it's leaping out of the chat box to literally click around your screen, you know, moving your mouse and typing in your apps.

0:23But before you fire your interns, is this actually saving you any money?

0:28Welcome to the deep dive.

0:29Right, because whether you run a small business or like an engineering team, you really need to hear this.

0:34We are pulling from the agent report, Flamerix AI, linear digressions, and mathematician Terry Tao today.

0:41Yeah, we're trying to figure out the true hidden costs and just the massive shifts of this new AI agent economy.

0:48Let's start with the big launches. I mean, OpenAI just dropped GPT 6.1 Tera at, what, an absurdly low $1 per million input tokens?

0:57Exactly a dollar. And Anthropic released Clawd's computer use and skills API at the same time.

1:04Right. So we have the capability and the token price is basically zero now.

1:09But the agent report and Flamerix AI are warning everyone not to take the bait on that token cost.

1:15It's a vanity metric.

1:17Wait, really? How so?

1:19Because the only number that matters now is the cost of a successfully completed task.

1:24Flamerix is running these $10,000 pilot programs right now, and they are ignoring API costs entirely.

1:29Oh, wow. So what are they tracking instead?

1:32They strictly measure how many human minutes of rework it takes to fix an agent's output.

1:37Which, I mean, that makes total sense.

1:39It like hiring a really cheap intern who takes like 50 tries to format a spreadsheet Yes exactly Their hourly rate is super low but the total cost of your time reviewing it and fixing all their mistakes and so on

2:10Figure out where they are.

2:11Right, taking and analyzing screenshots constantly.

2:15PM Lead ProFor a team lead: the hosts say cost per successfully completed task matters, not token price. Flamerix measures human minutes of rework on agent output in its pilots. Worth tracking on your own team.

2:28And linear digressions pointed out that the only way AI companies make this affordable is through heavy discounting on cached inputs.

2:34I think they said it makes them like 10 times cheaper.

2:37Up to 10 times, yeah.

2:38They store your session's visual context in the data center's KV cache.

2:43But the catch is that the infrastructure is optimized for continuous, unbroken processing.

2:49So if you pause a task to, like, grab a sandwich?

2:53Someone else's request comes in, and your context gets evicted from the cache.

2:58Wait, so I'm basically being financially penalized for taking a lunch break?

3:01Pretty much. If you step away, you have to pay cold-read premium prices to reload all that visual context when you get back.

3:09The system just punishes friction.

3:11So to survive this whole KV cache game your agents need to execute perfectly in one unbroken chain before your cache clears out which means they need flawless instructions on the first try Right And that where Anthropic new open standard comes in agent skills which GitHub Copilot just adopted actually Yeah instead of hoping the AI magically guesses your company specific legal review process from some long prompt agent skills packages that knowledge into version folders

3:40So the agent just dynamically fetches the exact domain expertise it needs right when it needs it. It turns abstract reasoning into a hard-coded, repeatable workflow.

3:49So it doesn't waste time or tokens stumbling around your desktop making mistakes.

3:55Exactly. It executes immediately, which keeps you within that cheap cash window.

3:59Okay, so we've basically solved the workflow problem for office chores, but if agents can

4:05autonomously execute procedural tasks, what happens when they scale up to cognitive automation?

4:12PM Lead ProFor a company owner: the hosts say agents should run in one unbroken chain to stay in the cheap cache window. Agent Skills package company-specific expertise, such as a legal review process, into reusable folders.

4:26When they start doing the core thinking, you mean. That's exactly what Terry Tao was warning about.

4:30He recently wrote this essay on the collapse of the theorem economy.

4:35Because traditionally, academia rewards mathematicians for the sheer manual labor of brute force and complex proofs, right?

4:42Right. But Tao argues that era is totally over.

4:45He says academia must start evaluating all math papers, assuming AI did the heavy computational lifting.

4:51Meaning the value shifts away from the actual mechanics of solving the problem.

4:55It shifts entirely toward why the question matters in the first place.

4:59You're rewarded for asking the right questions, yeah, not doing the manual labor.

5:03And if AI is automating high logic like that the underlying models have to become hyper Which we are seeing exactly in today model radar Like Qwen Dab and so on

5:35and Transformer.

5:36Which allows for perfect multi-line typography.

5:39We're also seeing deep local specialization with Google's Med Gemma.

5:43It's an edge-friendly model, fine-tuned strictly for local medical imaging and clinical text.

5:49So it doesn't need to know how to write emails, it just needs to read x-rays locally.

5:53Exactly.

5:54Same with JEV 1.14, this emerging decision class model.

5:58It's currently topping the AI index for pure logic routing.

6:02It doesn't generate text at all.

6:03It just makes decisions and routes tasks.

6:06So what's the big takeaway for the listener today?

6:09Stop tracking API token costs.

6:11Starting today, you need to track the exact number of human minutes required to review and fix an AI's output.

6:18That is your true cost of automation.

6:21I love that.

6:22Yep.

6:22And I'll leave you with this.

6:24If AI is collapsing the theorem economy in advanced math by automating the solutions,

6:30which of your company's core skills is actually just a soon-to-be-automated task economy?

6:36Think about that the next time you watch an agent clicking around your screen.

6:39PM Lead ProFind the transcript and sources on the episode page at pmlead.pro. Question for you: which of your team's core skills is really a task that could soon be automated?