Computational cognition

Eleven chapters, eleven levers. Architecture, scale, pretraining efficiency, alignment, democratization, context, test-time compute, RL at scale, inference systems, agency, and measurement - each one a separate bet with its own paper trail. The system you woke up to in September 2026 is what you get when all eleven are on at once. This chapter pins them on one map, closes the recurring threads (CoT, democratization, multimodality), names the villain that haunts every lever (Goodhart), and uses the map to design the small tool that has been waiting since Chapter 4: rewrite arbitrary text in your voice, without training anything.

11.1The eleven levers on one page

You can describe a modern model two ways. One: a 2017 architecture (Vaswani et al.) trained at scale on the internet with a 2024-vintage data pipeline and 2025-vintage numerics, post-trained for helpfulness and then for verifiable reasoning, prompted carefully, given room to think, served over an inference system that makes long decodes affordable, handed tools, and continuously measured. Two: a stack of optimizations applied to a frozen substrate - some baked into weights, some pinned into context. The through-line of the whole course is that the optimization moved from one place to the other as we figured out how, and the modern system uses both at once.

Toggle the levers and watch a (deliberately illustrative, not benchmark-accurate) capability bar assemble. The order is the order the field discovered them; it is not a strict dependency graph. Architecture without scale gets you BERT-era models that can read but not converse. Scale without alignment gets you GPT-3, which can recite anything and follow nothing. A modern pretraining recipe without a reasoning post-train gets you a fluent generalist that stops short of hard math. Add the rest and you get the assistant you used this morning.

11.2Three threads close here

The CoT seed grows three times

A single idea - make the model show its work - appeared at three levels of the stack. First in Chapter 3 as a prompt trick: write "let's think step by step" and accuracy on multi-step problems jumps without touching a weight (Wei et al., 2022). Then in Chapter 5 as a target: GEPA and MIPROv2 jointly optimize instructions and few-shot demos, so the search itself rewrites the chain of thought it asks the model to produce (Agrawal et al., 2026 ; MIPROv2 docs). Then in Chapter 6 as a weight update: DeepSeek-R1-Zero was trained with pure RL (GRPO + RLVR) on verifiable math, and a coherent reasoning trace emerged as the policy. On AIME 2024, R1-Zero went from 15.6% to 71.0% pass@1 during training, with 86.7% at majority-vote-of-64, "matching the performance of OpenAI-o1-0912" (DeepSeek-AI 2025).

Same idea, three levels: prompt → optimization target → weights. That is the whole point of the through-line. Once a useful behavior shows up at one level, the field finds a way to harvest it at every level.

The democratization curve

Every breakthrough got cheaper and more open. The alignment recipe shrank from RLHF's three-stage pipeline to DPO's "your LM is secretly a reward model" (Rafailov 2023) and then to GRPO's group-relative advantage with no critic at all (Shao 2024), before the field looped back to on-policy RL once verifiable rewards paid off (Chapter 7). Adaptation shrank from full fine-tuning to LoRA and then to QLoRA's 4-bit NF4 trick, which fits a 65B model on a single 48GB GPU (Dettmers 2023). The model layer shrank from closed APIs to LLaMA to DeepSeek-R1 to Qwen3 and Kimi K2, and by mid-2026 to DeepSeek V4.1, Kimi K3, and Qwen3.8 - open-weight flagships shipping within a season of the closed frontier.

The pattern is not just "things get cheaper." The pattern is that each generation's frontier trick becomes the next generation's default, and the new frontier moves up.

The eval shadow

How we measure is its own history. BBH set the scoreboard for prompted reasoning (Suzgun 2022). MMLU framed broad knowledge (Hendrycks 2020). SWE-bench moved the goalposts from quiz answers to real software work (Jimenez 2024). Then agents raced past every fixed target and the field needed new axes: task-length horizons from METR, a rotating set of contamination-hard evals, and interpretability tools that watch model internals directly. Chapter 10 tables the 2026 landscape. It is the chapter that keeps this course honest.

11.3Multimodality, the thread that opened in Ch1

When you read about ViT and CLIP in Chapter 1, the point was that vision joined the same architectural family. That thread closes here. By 2024, frontier models were natively multimodal across text, image, audio, and (with longer contexts) video: GPT-4o shipped May 13, 2024 with text/image/audio in and out; Gemini 1.5 Pro was announced Feb 15, 2024 as "natively multimodal" with a 1M-token context preview (tech report).

Through 2025 the sub-modality frontier moved on its own. OpenAI's native image generation replaced DALL-E inside 4o on March 25, 2025 - autoregressive image tokens produced by the same model that produces text (OpenAI). Google's Gemini 2.5 Flash Image ("Nano Banana") shipped Aug 26, 2025 with character consistency and multi-image blending (Google); the Gemini 3 Pro Image tier ("Nano Banana Pro") followed on Nov 20, 2025, two days after the Gemini 3 launch. Video followed: Veo 3 (Google I/O, May 20, 2025) generates synchronized audio with the video track (DeepMind); Sora 2 (Sep 30, 2025) improved physics and synchronized audio at OpenAI (OpenAI). "Natively multimodal" is now table stakes; the interesting frontier is what each additional modality lets an agent do.

11.4The capstone: a voice tool, finally placed on the map

Back in Chapter 4, when we walked through QLoRA, there was a personal note. The reader had been carrying around a 2019-era idea: fine-tune a small model on their own writing so it could rewrite arbitrary text in their voice. Chapter 4 placed QLoRA as a real and valid democratization node - that idea was not wrong, exactly. It was simply the right idea for the wrong lever. The map we just drew tells us which lever it should be.

Style is steering, not knowledge. Voice is a direction in a frontier model's existing capability surface, not a new fact set you need to bake into weights. That is precisely the regime where discrete prompt optimization shines, and it is what makes the prompt lane (Ch 5) the correct lever here.

Three things decide it, and all three favor a frozen frontier model:

• Quality ceiling. A frontier model already writes better than any small model you could realistically fine-tune. You only need to steer it.
• Iteration speed. "Sounds like me" is subjective and needs a human in the loop. Prompt search iterates in seconds; fine-tuning iterates in hours.
• The artifact. The output is a prompt you can read, edit, and carry to any model, not a LoRA welded to one base.

The recommended build, in primitives the literature already verified

Nothing in the build below is new science. It is a composition of pieces every previous chapter introduced: GEPA's reflective mutation and Pareto frontier (Ch 5), DRPO's tuning-free prompt search (Ch 5), a calibrated LLM judge (see Ch 10 for the general treatment of judges and their biases), and Chatbot Arena's pairwise-to-Elo aggregation. The novelty is in the recipe and in naming the failure mode.

That fitness signal is the only genuinely new component, so it is worth feeling. Click which rewrite sounds more like a single coherent voice; an Elo ranking emerges from nothing but binary choices. This is the loop running inside step 4 of the build.

The villain, named

The same Goodhart attractor from earlier returns here as judge-drift. If the LLM judge is the loop's fitness function, the search will eventually find prompts that please the judge in ways the user would not endorse. The literature already gives three mitigations and they all stack:

• Pareto selection over multiple objectives, so the search cannot collapse to one. GEPA does this natively.
• Periodic human-pairwise re-calibration of the judge, the same loop DRPO and Chatbot Arena use to keep their signals honest. Chapter 10 lists the ways this loop itself can be gamed, and how leaderboards responded (style-controlled Elo; the Leaderboard Illusion).
• Held-out audit prompts the judge never sees during search, so reward-hacking is detectable instead of invisible.

What this resolves

The reader's original instinct - train a small model on my writing - was not wrong. It was a real and valid Chapter 4 node, and the QLoRA paper made it practical for someone with one consumer GPU. But the convergence the course has been tracking changes the answer. Optimizing the context can now rival optimizing the weights, far cheaper, with an artifact you can read, port, and audit. The voice tool is the same idea as the reader's 2019-era plan, executed on the lever the field has spent four years sharpening into the better one.