中文
Cursor Compile

Entropy-reduction physics explains why token maxing keeps costing more for less: context is the other half of efficiency the whole industry skipped.

Intelligence Efficiency, Ben Geist | Compile 26 · Ben Geist

15 min
ContextAgent

15 min total·Actually worth watching closely: ~5 min·3 must-watch clips

Orange = the 5 minutes worth watchingFor the rest, the guide is enough
Segment guide · 7 segments
  1. 0:00 1:45Skim

    You're buying tokens, not intelligence

    Opens with the core contradiction: Ramp's data shows monthly token spend up 13x since January 2025, but returns on intelligence follow a logarithmic curve, and Uber and Meta have already started clamping down on token maxing.

    Paying for intelligence is really paying for tokens; the two are not equal, and exponential consumption isn't economically sustainable

    The spend-versus-intelligence curves (around 48s and 70s) are worth a glance; the rest is spoken setup▶ Jump to 0:00
    Speaker · Ben Geist
  2. 1:45 4:10Skim

    The LLM as an entropy reduction machine

    Combines Shannon entropy with the thermodynamic view: how far a system's entropy can be reduced is bounded both by the work you put in and by the information you already have — like cleaning your room, you have to know where the dirty clothes are before the work pays off.

    Entropy reduction has two levers, work and information; the chart at 163s names the goal — bend the intelligence curve from logarithmic toward linear

    The entropy frame is the key to following the second half; the chart at 163s is worth watching, and the equations can be skimmed off the slides▶ Jump to 1:45
    Speaker · Ben Geist
  3. 4:10 6:11Listen

    The industry bet on only one half: work

    Points out that today's advances — chain of thought, reasoning — sit almost entirely on the "more work" side, while the information side's mainstream tools, RAG and memory, don't scale well, are used for needle-in-a-haystack semantic search, and stay stuck in token space rather than latent space.

    The efficiency bottleneck is the neglected information/context half, and existing methods never reach latent space

    Pure argument with no key visuals — fine to listen to like a podcast▶ Jump to 4:10
    Speaker · Ben Geist
  4. 6:14 9:20Watch

    Experiment 1: a shared KV cache across agents

    The classic supervisor-worker architecture rediscovers the same context over and over and burns an astronomical number of tokens; the fix is a globally shared KV cache, with each newly spawned worker initialized from task-relevant context compressed out of it.

    At unchanged accuracy, worker tokens drop 42–57% and total tokens 21–31%

    The architecture comparison drawn from 396s is the key to seeing where the waste is and how the cache is shared — visually dense▶ Jump to 6:14
    Speaker · Ben Geist
  5. 9:20 11:20Skim

    Experiment 2: sparse attention as a reranker

    Models sparse attention as a reranker: document scores computed from the queries and keys of DeepSeek's lightning indexer (ReLU of the normed cosine scores summed over heads, averaged over tokens within a document), benchmarked on a multi-hop question-answer dataset.

    Sparse attention matches or beats SOTA reranker models — its efficiency comes down to accurate context retrieval over its own token space

    The scoring equations and comparison table at 595s are worth pausing on; the narration just follows the slides▶ Jump to 9:20
    Speaker · Ben Geist
  6. 11:20 13:17Listen

    Experiment 3: memory as its own modality

    Work with Stanford's SNAP lab treating context injection as a modality, the way text, image and video are: a memory module takes docs plus a query and compresses them into 16 latent representations injected straight into a frozen Qwen 8B, with no change to the base model.

    Context can bypass token space entirely and enter a frozen LLM as latent representations — the most radical rework of the information side

    Mostly spoken method description; catching the pipeline is enough, and the results chart comes at the next stop▶ Jump to 11:20
    Speaker · Ben Geist
  7. 13:17 14:34Watch

    372x compression and three criteria for adoption

    The memory module beats RAG-50's 55% exact match on TriviaQA while using 372 times fewer representations in the input; the close proposes three criteria for an ideal context-injection model: low latency, scaling over large corpuses, and essentially zero switching costs to the base model.

    The three criteria are a selection checklist: usable at inference time, scalable, and reusable when the next base model ships without retraining

    The results chart at 798s carries the talk's conclusion, and the three-criteria slide is worth a screenshot▶ Jump to 13:17
    Speaker · Ben Geist