These startups are chasing the next big thing in LLMs
View original at technologyreview.comMIT Technology Review - Ai Research Title: These startups are chasing the next big thing in LLMs Date: 2026-08-10 09:00 Source: https://www.technologyreview.com/2026/08/10/1141511/these-startups-are-chasing-the-next-big-thing-in-llms/ <div></div> <p>MIT Technology Review<em>’s What’s Next series looks across industries…
What we drew from this source
The claims Via News extracted from this document. We point to the source; we don't replace it.
The human brain is an AGI system that operates on 20 watts of power, suggesting AI can become far more efficient and innovative.
60% confidenceIt is very difficult to represent a sudoku board word by word, illustrating a limitation of language-based reasoning in transformers.
60% confidenceDiffusion LLMs still use a big transformer model but can predict many tokens at once, making them faster and more cost-efficient than most competitors.
60% confidenceInception is bullish about the diffusion approach because it is the one that will scale up.
60% confidenceLiquid AI's designer AI, which selects combinations of neural network types, is the core technology of the company.
60% confidenceThe ultimate goal for AI is not solving puzzles like sudoku but tackling problems with no existing playbook, such as curing cancer.
60% confidenceUnlike image pixels, there is no intermediate representation between discrete words like 'cat' and 'dog', which made adapting diffusion to text a challenge.
60% confidenceGoogle's competing diffusion effort is validating for Inception's approach.
60% confidenceTransformers underpin the entire AI industry and are among the most important innovations in the history of computer science.
60% confidencePower retention has many useful applications, from analyzing hours-long videos to building agents that can stay on task for weeks.
60% confidenceHuman insight ('eureka moments') is not necessarily expressed in language, so reasoning constrained to language is limited.
60% confidenceTransformers were an engineering convenience rather than an inevitable endpoint, and further breakthroughs beyond them are likely.
60% confidenceThe only things that ultimately matter for LLMs are speed and cost, i.e. intelligence per dollar.
60% confidence
Data points we hold from this source
| OpenAI · computing spend | 50 USD |