Article: https://www.interconnects.ai/p/deep-research-information-vs-insight-in-science

(sorry about some more audible breaths in this -- I'm going to work on it!)

We at Ai2 released a local LM iPhone app for our OLMoE model (1B active, 7B total params), with greatly improved scores! Let us know what you think, or read more here.

OpenAI’s Deep Research has largely been accepted as a super valuable tool for knowledge workers and analysts across the economy, but its real engine of economic progress is going to be changing the nature of scientific progress. Science is the fuel of technological revolutions.

Deep Research in its current form feels like a beta version of a next-generation piece of technology. It does what it is tasked with — searches the web and processes many resources to create a useful report with referenced sources. Some of my uses include researching model evaluations, recent robotic learning research, and AI for science breakthroughs.

Deep Research’s limitations mostly feel like problems of search, where it is prone to returning SEO optimized slop, style, where it returns verbose, low information density writing, and modality, where it does not have the ability to read, process, and return plots and diagrams. All of these are surely solvable and expected features if we look at the rollouts of other AI models in the last few years.

This isn’t a product review (you can read Stratechery or Turing Post for more of that) — as the answer is quite simple, if you work in a knowledge intensive vocation you should be using this — but rather asking: So what comes next?

The place to start from within AI circles is to revisit the question of “When will AI make novel discoveries?” A good example of this is in the Dwarkesh Podcast episode with Dario Amodei:

One question I had for you while we were talking about the intelligence stuff was, as a scientist yourself, what do you make of the fact that these things have basically the entire corpus of human knowledge memorized and they haven't been able to make a single new connection that has led to a discovery?

An example experiment we could do to test this is to train models on time-gated information and see if it can repeat a scientific discovery we already made (yes, this would be difficult to run, but not impossible). Ross Taylor described this on his Interconnects Interview:

So an experiment I've never done because I didn't have [the] compute would be this. Imagine if you could train a language model on all documents up to 1905, which is the year when Einstein had his miraculous year of four seminal papers. With that model, which is trained up to 1905, could you prompt the model to come up with a good explanation of the photoelectric effect, special relativity, this kind of stuff? And what would it take to rediscover these things?

The dream is for AI to make breakthroughs, and the absence of evidence for this even after the release of Deep Research is driving a reckoning over what language models will ever be able to do. The fork in the road is either believing that scaling (either in parameters or in new training methods) will unlock “insights” or accepting that the current generation of models are very useful tools and nothing more supernatural. Likely the most powerful tool humanity has made yet. Our first power tool for information.

Much of science is not about making novel insights but about making progress within established problems of the field. In AI, these are the countless benchmarks we are saturating. A very valuable contribution in AI as a field can be re-using known resources in a simpler way.

With AI, we are going to learn the boundary between true insight and scientific progress. A related form of scientific progress is the compression of noisy ideas and experiments into a cohesive trend. Something that Deep Research can likely do, but not something that builds the allure of Einstein and the other scientific greats.

To understand this relationship between Deep Research, AI broadly, and the nature of science, we must address:

* How to interpret existing “AI for Science” projects like AlphaFold in the bigger context of science,

* How reasoning models, AI research systems like Deep Research, and other forthcoming AIs revolutionize existing scientific practices,

* How recent developments in AI challenge Kuhn’s formulation of scientific revolutions, and

* How current institutions will need to change forever in the face of AI?

This (hopefully) series of posts is my attempt to create a worldview around what science means in the face of AI. Today, we focus on the first two — major AI for science projects and how normal science is being accelerated by AI — and hopefully raise urgency within the community to consider the final question.

The starting point — grand AI for science projects

There is a substantial overhang in computational infrastructure and fundamental deep learning capabilities relative to their impact on the broad class of sciences. In order to make a substantial leap in the application of AI to a specific domain, a team must mold the existing underlying capability of AI to the needs of trained groups of scientists.

The list of examples people think of in this mold ranges across domains: AlphaFold for protein folding, AlphaGeometry for mathematics, GraphCast and GenCast for weather, and more that lack such prominent branding. They leverage advancements in deep learning and transformer architectures, but tend to have X-factors specific to the domain of interest (see a Deep Research query summarizing this). Such added features are pulling forward AI capabilities to suit a narrow domain.

There’s a substantial craft to selecting suitable problems for applying this grand AI for science approach. It requires a field with central elements that are quantitatively focused. Even with this, outcomes are more uncertain than standard AI research or standard research in the domain of choice.

The essay A new golden age of discovery from AI Policy Perspectives details how DeepMind sees the opportunity here and showcases some internal ingredients they found that make these projects more likely to be successful.

The fact that any of these projects have succeeded shows the overall potential of AI for science. The overall necessity of the approach depends on whether the grand AI for science breakthroughs are pulling forward progress by months or years, or if these models are the single required breakthrough to approach entirely new areas of study.

As the broader scientific community embraces AI as “something that works” more of these step changes will happen. They take a very large density of compute and talent on a single problem.

These projects fit more naturally into a classical view of science. They take substantial resources and are high risk. Meanwhile, the mass market AI tools that everyone is adopting will dramatically shift the practice of doing science.

Towards instantaneous Ph.D.’s

We have two tools that dramatically shift the nature of scientific exploration. They will only get better.

* AI models that excel at code, mathematics, and reasoning: OpenAI’s o3, DeepSeek R1, Gemini Deep Thinking, etc.

* AI systems to rapidly parse and summarize existing literature: OpenAI’s Deep Research, Gemini Deep Research, Ai2’s Scholar QA (specific to academic papers), and many more that will come soon.

These tools are dramatically accelerating the most time-consuming aspects of research, particularly in computationally intensiv