The fact that LLMs lack a constrained inner monologue feels like a fairly catastrophic weakness that needs to be changed. They're certainly given time to think between tokens, but that thinking is fixed compute, brief, and left totally undiscretised.

NVIDIA’s market cap has increased by ~the GDP of New Zealand since yesterday’s earnings announcement. We are currently going through the single greatest technological shift in human history, so this checks out. Autonomy, longer context, and embodiment are all inbound.

High stakes AI customer service as an API with an intuitive dashboard, if executed properly, would be a billion dollar product. Needs low latency response times on calls, natural sounding voice/language, adaptable, knows when to triage to human assistant, etc.

Foundation models will become modules of more elaborate systems and networks. Councils of models will reach agreement by majority vote on high stakes decisions, specialist fine tuned models will divide and conquer work orchestrated by an overseer model, etc. We need compute.

Cityscape by Tomoaki Okada

Tsarathustra A vast amount of undiscovered knowledge is latent via induction from nodes of existing knowledge. Foundation models will be the first entities that are simultaneous experts in thousands of fields. Interdisciplinary discoveries should be quite prevalent with sufficient compute.

Claude 3 was trained on synthetic data (“data we generate internally”). Fairly clear that compute is the bottleneck given that parameter count and data can be scaled.

finbarr The technical overhang is compounding from multiple angles - exponentially growing context windows with high accuracy retrieval, improvements in quality of token outputs, increased modalities, external tool use, long term memory, synthetic data, compute improvements. Dizzying.

Garry Tan All organizations are going to become orders of magnitude more efficient with AI. The primary hurdles to transitioning to nearly/fully automated institutions will be social inertia and politics.

Teortaxes▶️ This feels like a trillion dollar discovery. Does this not completely upend NVIDIA? (See suggestion of “new hardware” in screenshot)

Deedy Moats washing away like sand castles, yet the capital flows. Can’t think of another sector with such a lack of lasting power for a given product that gets this amount of funding.

Pata van Goon Follow the flops.

“Self play allows you to turn compute into data.” - Ilya Sutskever, 2018 Just after this portion of the lecture, he discusses agents. It is clear that autonomy is the next step. Life 2.0 will be a compute-bound phenomenon. ASML and TSMC are the midwives.

According to Daniel Gross’s interview of Eric Steinberger (Magic AI) they have a model with at least a 100M token context window, if not far greater. Attention, despite quadratic scaling, is continuing on an exponential. A 12,500x increase from GPT-4’s 8k window a year ago.

Meta is currently training a 405B model that is at 85% (and climbing) on MMLU.

Sheel Mohnot The US is swiftly facing the music - that decades of offshoring has rendered their command over material reality inferior to that of China’s. It’ll take decades to reverse this - likely via robotics and a high degree of process automation.

Scaled inference-time compute yields significant improvements to output quality and is likely the basis of Open AI's upcoming strawberry release. The rate of economic progress is going to increasingly correlate to the volume of installed FLOPs.

Scale inference-time compute in verifiable domains to yield synthetic data, distill frontier model trained on synthetic data to lean model, commercialize. Repeat.

Andrew Curran As Larry Ellison said, we will need “acres of NVIDIA GPUs” to find out There comes a point where the flywheel begins to drive itself. The installed compute will continuously become more optimally allocated with less and less human input. Economic progress as a function of FLOPs

Leopold Aschenbrenner’s Situational Awareness made what, in June, were extraordinary claims of trillion dollar infra buildouts to support the energy and compute needs to achieve AGI/ASI. As the haze clears, it appears that he was spot on - there is no precedent for the stakes involved.