AI - 2024AD: 212-page Report (from this morning) Fully Read w/ Highlights

A six-year “State of AI” report released by Andreessen Horowitz (a16z) Capital frames 2024 as a year when leading models stopped feeling like separate species and started converging—while costs, compute, and multimodal capability kept accelerating. The report’s headline theme is that major systems are increasingly overlapping in what they can do because they’re trained on similar large-scale data, pushing models such as Claude 3.5, Sonnet, Grok 2, and Gemini 1.5 toward GPT-4–level behavior rather than diverging into fundamentally different approaches. That convergence matters because it shifts the competitive question from “who has the best architecture” to “who can scale reliably, ship safely, and monetize faster.”

The report also revisits earlier forecasts and measures them against reality. One prediction—about spending more than a billion dollars to train a single frontier model—was judged harshly at the time, but later reporting suggests the scale is now firmly in the billions. OpenAI’s projected 2024 training costs were cited as roughly $3 billion for frontier-model training, excluding additional compute used for research and iteration. The same reporting also pointed to a longer runway for profitability, with internal sources suggesting OpenAI may not turn a profit until 2029. The implication is that the frontier race is not just about breakthroughs; it’s about sustained, expensive compute cycles.

Multimodality is treated as a practical turning point rather than a novelty. Meta’s Movie Gen is highlighted for generating audio alongside video, and the transcript contrasts that paper-based progress with tools that let users create short clips immediately—upload an image, choose an effect, and generate “melt,” “explode,” or “squish” style transformations. The report’s broader point is that models are moving from single-task text generation into systems that can manipulate multiple media streams in one workflow.

Science and medicine are also pulled into the spotlight. The Nobel Prize in Physics and Chemistry is mentioned in connection with neural-network-driven work (including AlphaFold), reinforcing a narrative that AI is increasingly “eating science” by accelerating discovery. One standout example from the report is “brain LM,” a transformer-based model that can be fine-tuned to predict clinical variables such as age and anxiety disorders from brain activity, and potentially support in-silico medication testing by simulating biologically meaningful responses.

Hardware and scaling dynamics get their own section. A chart tracks how quickly Nvidia data center GPUs are released and how teraflops per chip rise over time, while also noting that clustering more GPUs across data centers multiplies the effective compute. The transcript ties this to a belief that the next two years of progress are largely “baked in” by sheer scale.

Yet the report doesn’t treat risks as solved. Jailbreaking remains an active problem, with stealth attacks and compromised instruction hierarchies reported to persist for hours. A taxonomy of real-world misuse is used to ground concerns in current harms—especially impersonation scams and non-consensual intimate image generation—rather than hypothetical future catastrophes.

Finally, the report’s next-year predictions are presented as a mix of firm and vague claims. The transcript’s own critique focuses on how hard it is to measure “meaningful changes” or “breakout status” without clear metrics. It ends with a forward-looking bet: that the next major leap will come from models that break state-of-the-art performance across multiple modalities at once, alongside continued rapid scaling and valuation growth.