(Spine Ch. 1.)
“By now, ‘GPS’ is a colorless term denoting a particularly stupid program to solve puzzles. But it originally meant ‘General Problem Solver’, which caused everybody a lot of needless excitement and distraction. It should have been called LFGNS: ‘Local-Feature-Guided Network Searcher’.” Drew McDermott, Artificial Intelligence Meets Natural Stupidity, SIGART Newsletter No. 57 (1976)
Two seconds of latency, two cents a ticket, and a $50K/month bill sitting on a CFO’s desk. The CTO wants to wait for quantum. The CEO just read a piece saying the bubble is about to pop and wants to wait for that. The CFO stops waiting and freezes the whole project on a Tuesday. Nobody in that room is stupid. Everybody in that room is about to make the same call three different generations of very smart people already made, in 1974, in 1987, and in 1993, and I have personally made a dumber version of it with my own money. The bill was not the problem.
Bottom line: AI is not new. It’s Turing asking “can machines think?” in 1950, machine learning in the 70s, backprop in the 80s, and LLMs in the 2020s, the same loop each time, waiting for hardware and data to catch up to the theory. My read: the next likely move is ASICs for local inference, then, if the bet pays, quantum. Know the arc and you stop confusing “new to your feed” with “new,” and you stop betting the company on the next demo before the last one finishes delivering.
1950: Turing publishes “Computing Machinery and Intelligence.” Can machines think? He proposes the Imitation Game: if a human can’t tell the difference, it works. Seventy-five years later, that’s still the best definition we have.
1956: Dartmouth Summer. McCarthy coins “artificial intelligence.” Ten guys promise thinking machines in a summer. It was adorable and wrong by about twenty years. This becomes tradition: promise general intelligence, deliver a demo that works on one narrow task, lose the money, keep the grad students.
1966: Weizenbaum builds ELIZA. Pattern matching, a simulated therapist that repeats you back with a question mark. People bond with it. Cry on it. Mistake it for understanding. This is the first warning about Ch. 5: humans will fall in love with a sufficiently confident parrot. We ignored this for 60 years. I ignored it for about 60 seconds, which is how long it took me to start thanking one of my own agents in the prompt, in production, on a loop that ran a few thousand times a day. I was paying tokens to be polite to a for-loop.
The winters (1970s, 1980s, 1990s): Every time the demo outruns the hardware, funding dies. Lisp machines. Expert systems. Nets that needed a year to train 10K examples. Each winter culls the tourists. The grad students keep working because they have no choice.
1986: “Learning Representations by Back-propagating Errors,” Rumelhart, Hinton and Williams, Nature, 1986. Nets can learn multiple layers. Nobody outside the lab cares: no data, no compute. I spent years calling it “Hinton’s paper,” which is a great way to erase two co-authors in three words.
1997: Two things. Deep Blue beats Kasparov (brute force, not thinking, but it moves the goalpost). Hochreiter and Schmidhuber publish “Long Short-Term Memory” in Neural Computation, giving nets a memory for sequences. Both filed under “neat.” Neither one ships into anything you would use. The internet hasn’t dumped the world’s text into a bucket yet.
2012-2017: Deep learning starts winning competitions. ImageNet. Speech. Then “Attention Is All You Need,” the transformer paper. The paper is not the revolution. The revolution is what happens when you feed it petabytes and GPU clusters. The paper was the permission slip.
2020-2023: GPT-3 (expensive, API only). Then ChatGPT (free, text box, and your mom can use it). The interface mattered more than the parameter count.
2023-2026: Claude, open weights (Llama, Mistral), context windows hit 100K, function calling works, and agents that loop actually loop. The model stopped being the product. The brain got hands. That’s why this book exists now.
The last three years were about giving the brain a job. Loops. Tools. Checkpoints. Stop conditions. That’s not chemistry, that’s plumbing. It is also considerably less fun to put on a slide than “we are building a brain,” which is exactly why I spent a year building the brain and about a week on the stop condition, and then acted surprised when the thing would not stop. The loop kept going the way AOL kept mailing out free CDs, cheerful and relentless and impossible to make stop, except I was the one who had addressed the envelopes. Somewhere around iteration nine hundred I stopped hearing “you’ve got mail” and started hearing a smoke alarm.
Move 1: ASICs everywhere. Right now you’re renting inference from the cloud. Azure, OpenAI, Google. Each call home costs money and latency. The next evolution is ASICs that run models faster and cheaper than anyone can rent them: your box, your phone, your edge device. Groq has been building it since 2016, when a group of Google TPU engineers walked out to make a chip that does one thing: run inference. They call it an LPU. GroqCloud opened the doors on February 19, 2024. A $640M Series D in August 2024 valued the company at $2.8B. Then in December 2025 Nvidia licensed the technology, in a deal press coverage put in the tens of billions. I have not read the filing myself, so take the size as reported and not as checked. The direction tells you the whole story in one number: the crew that left to undercut Nvidia ended up renting Nvidia the shovel. When that lands at scale the economics invert, and it’s cheaper to run local inference a hundred times than to pay for one API call. That breaks the cloud-rental model and kills the “AI is expensive” narrative. If everyone has the same model running on the same cheap silicon, the edge belongs to whoever built the best agent loop around it.
Move 2: Quantum. This is the one people lose their minds over. Quantum computers can solve certain classes of problems (factoring, optimization, simulation) exponentially faster than classical computers. But quantum isn’t “AI but faster.”
Here’s my read. A quantum machine works by arranging amplitudes so the wrong answers cancel and the right ones reinforce, which only pays off on problem classes shaped to allow it: factoring, optimization, simulation. Nothing about that makes LLM inference faster today, and nobody has shown that it will. Which makes stop conditions more important, not less: whatever fan-out you eventually get is still a loop, and a loop that never quits isn’t a breakthrough, it’s an outage with better branding. The chips are not vapor. Google announced Willow on 9 December 2024, 105 qubits, and error correction that improved as the chip grew, a first. Microsoft followed with Majorana 1 on 19 February 2025. Neither one runs a language model today. “Shortly” is a word I have retired.
When people promise it works before it does, it’s a winter waiting to happen. The 2026 version of “AI will be AGI by 2030” is “quantum will make AGI trivial by 2035.” I once reorganized a roadmap around a technology I could not have explained to a bored teenager. If you can’t draw it on a napkin, you are not betting on it, you are buying a lottery ticket with a quarterly plan stapled to it.
This loop has run at least three full turns since 1950 (1974, 1987, 1993). It will run again. The people who know the cycle don’t confuse themselves with the hype. The people who don’t know it lose the company.
These numbers are from memory, no invoice, so treat the shape and not the digits. 2024: a support agent prototyped on OpenAI’s API, drafting tickets, classifying severity, escalating. Roughly two cents a ticket, roughly two seconds of latency, roughly $50K a month. Same CTO, CEO, and CFO reactions as the opening.
What I remember of the fix: fine-tune a small open-weight model on about 10K of your actual tickets, deploy it on-prem or on an edge cluster, and the per-ticket cost lands roughly two orders of magnitude cheaper, well under half a second. The artifact that would settle it: a provider billing export next to an on-prem GPU-hour bill for the same month. I remember it as a small open-weight model and about ten thousand labeled tickets. I never pulled the invoice or the repo, so treat it as a war story, not a benchmark. The unglamorous part is that the whole win lived in boring tickets nobody wanted to label, which is exactly why I bought the expensive branded version instead.
The loud failure is betting on hype: “AI will solve it” without building the loop. Easy to spot.
The quiet failure is worse:
You treat “new to my feed” as “new to the world” and you panic-build.
Someone on Twitter says quantum is here. You kill your ASIC roadmap. Someone says agents are overhyped. You freeze the agent project. You’re not building for the future, you’re pattern-matching to press releases. I have killed a working roadmap over a tweet with 400 likes. I have production systems with more daily users than that, and I let a number smaller than my error log set my quarter. I used to watch the Slashdot effect flatten a server in four minutes and know exactly what it meant. A tweet is not the Slashdot effect. It is one guy with a keyboard getting first post, and I rewrote a roadmap for him like a n00b.
Second quiet failure: you think the move is the tech, not the timing.
Transformers were brilliant. GPT-3 shipped in 2020 as a paid API almost nobody outside dev circles touched, and ChatGPT in 2022 made it usable by anyone with a browser. The people who understood the timing (not the paper) made the move.
Do
Don’t
This chapter is the throughline: AI is not new, the loop repeats, and the pieces are finally all there at once (models work, agents can loop). Ch. 2 is the money story. Ch. 3 is what actually works. Everything after is how to build loops that survive the cycles.
Thesis is Jeremy’s: AI’s a 75-year loop, next moves are ASIC + quantum.
Verified / load-bearing: - Turing (1950): “Computing Machinery and Intelligence.” Real paper, real argument. - Dartmouth (1956): McCarthy et al. Real event. Real miss on timeline. - ELIZA (1966): Weizenbaum. Real demo, real attachment. (Anchor for Ch. 5.) - Deep Blue (1997): documented. - Backpropagation (1986): Rumelhart, Hinton, Williams, “Learning Representations by Back-propagating Errors,” Nature, 1986. Hinton is a co-author, not the sole author. https://en.wikipedia.org/wiki/Backpropagation - LSTM (1997): Hochreiter and Schmidhuber, “Long Short-Term Memory,” Neural Computation 9(8): 1735-1780, 1997. https://en.wikipedia.org/wiki/Long_short-term_memory - Transformer paper (2017): “Attention Is All You Need,” Vaswani et al. - GPT-3 (2020), ChatGPT (2022): documented public releases. - Open weights (2023-2026): Llama, Mistral, others. - Function calling (2023+): OpenAI, Anthropic, others. Standard feature. - Agent loops (2023-2026): implemented by practitioners, me included. - Groq founding (2016), GroqCloud launch (Feb 19, 2024): Wikipedia, “Groq,” https://en.wikipedia.org/wiki/Groq - Groq $640M Series D at $2.8B valuation: Forbes, August 2024, cited in https://en.wikipedia.org/wiki/Groq - Nvidia licenses Groq technology, December 2025: reported by press at a tens-of-billions figure. Printed as reported, not as verified from a filing.
Stated as memory, not receipt: - Support-agent worked example (2024): the cents, seconds, monthly bill and the drop after fine-tuning are recollection, printed as such. No invoice or repo backs it.
Honest gaps: - No academic papers on “AI cycle hype”: observed pattern, not peer-reviewed. - Where deployed ends and bet begins: everything through 2023-2026 is shipped and checkable. Move 1 is shipped hardware with an unshipped economic conclusion. Move 2 is a bet, labeled as one. - Cross-refs kept honest: Ch. 2 (money), Ch. 3 (capability), Ch. 5 (attachment), Part II (loops).