Very soon, the blocker to using AI to accelerate science is not going to be the ability of AI, but rather the systems of science itself, as creaky as they are.

The scientific process is already breaking under a flood of human-created knowledge. How do we incorporate AI usefully? Science isn't just a thing that happens. We can have novel discoveries flowing from AI-human collaboration every day (and soon, AI-led science), and we really have not built the system to absorb those results and translate them into streams of inquiry and translations to practice

Some new theoretical economics papers looking at the implications of AGI.

These two papers argue that a true AGI-level AI (equivalent to a human genius), if achieved, would eventually displace most human labor and reduce the economic value of remaining human work to near-zero.
Papers:
conference.nber.org/conf_papers/f2…
conference.nber.org/conf_papers/f2…

Hey Claude, ChatGPT, Gemini: "I am time traveling back to the 75 BC Rome for one day. I can't bring anything back. What is the one thing I could learn that would most advance today's knowledge and what is one thing I could do there that would make me richest today"

Pretty good

Summary of their views:
Gemini: Get how to make maritime concrete and provide an artifact proving time travel
Claude: Memorize specific texts and the formula for concrete and location of proscribed villas
ChatGPT: Figure out Etruscan language and the location of Alexander's Tomb

Claude's new ability to work with Excel files is the best I have seen so far

I have given it existing spreadsheets to work with and asked it to create new ones. Good use of formatting, formulas, etc.

It created all of this, including 406 formulas, from one prompt (& its solid).


This is like an assignment I give, and it would be a good result of a week-long team project for my MBA class. I can't promise it is error free, but I haven't found any issues so far.

This chart is being horribly misinterpreted.

This is not where the training data of AI comes from, it is a study done by a SEO firm that claims to show how often sites come up at least once in THE WEB SEARCH FUNCTION of certain AI agents when they do a web search for more info. The company searched for a bunch of keywords using Google AI Mode and ChatGPT web search and Perplexity and then said they measured how many times these sites were included in the reply.

If you are search for "find me a good stove" or whatever, this should look like the results.

We have data on the environmental impact per AI prompt:
Gemini: 0.00024 kWh & 0.26 mL water
ChatGPT: 0.0003 kWh & 0.38 mL
...the same energy as one Google search in 2008 & 6 drops of water.

Seems to be improving, too: Google reports a 33x drop in energy use per prompt in a year.

These numbers match independent direct measures: 0.00004 kWh for 400 tokens on Llama 3.3 70B on a H100 node.

We do not know the amount of energy required to train these models, which was estimated at a little above 500,000 kWh for GPT-4, about 18 hours of a Boeing 737 in flight.

“Claude Opus 4.1, De-Carcinize the Great Gatsby” (That was the only prompt)

Pretty clever, actually.


# The Great Gatsby: De-Carcinized

## Chapter 1: Direct Approach

Nick Carraway moved to West Egg and immediately walked straight to his neighbor Jay Gatsby’s house.

“Hello,” Nick said. “I’m your new neighbor. I notice you throw parties.”

“Yes,” Gatsby replied, opening the door fully instead of peering through a crack. “I throw them hoping Daisy Buchanan will attend. She lives across the bay. I love her.”

“She’s my cousin,” Nick said. “She’s married to Tom.”

“I know. Would you arrange a meeting?”

“Sure.”

## Chapter 2: Honest Communications

Tom Buchanan stood in his living room, addressing everyone directly. “I’m having an affair with Myrtle Wilson. Daisy, you probably know this. Myrtle, you should know I’ll never leave Daisy for you.”

Myrtle nodded. “I understand but choose to continue anyway.”

Daisy turned to face them. “I’m unhappy but too comfortable to change anything.”

## Chapter 3: The Party Without Pretense

Gatsby’s next party featured a sign at the entrance: “These Parties Are Specifically to Attract Daisy Buchanan’s Attention.”

Guests arrived and immediately stated their intentions:
“I’m here for free champagne.”
“I’m networking for business.”
“I enjoy judging rich people.”

Gatsby stood on a platform with a megaphone: “Thank you all for coming to my Daisy-attraction event. Please enjoy yourselves while I wait hopefully by the window.”

## Chapter 4: Swimming Forward

Gatsby drove Nick directly to the city—no detours through the Valley of Ashes, no sideways routes through East Egg.

“I should tell you my backstory without any embellishment,” Gatsby said. “I was born poor as James Gatz. I made my fortune through questionable means with Meyer Wolfsheim. Everything I’ve built is to win Daisy back.”

“That seems unhealthy,” Nick observed.

“It absolutely is,” Gatsby agreed.

## Chapter 5: The Reunion

At Nick’s house, Gatsby walked straight to Daisy when she arrived.

“I’ve loved you for five years,” he said. “I’ve built my entire life around the hope of resuming our relationship.”

“That’s concerning,” Daisy replied. “But I’m flattered and bored in my marriage, so let’s proceed.”

They looked at each other directly, no dancing around the truth.

“This will end badly,” they said in unison.

“Want to see my house?” Gatsby asked.

“Yes, though material possessions won’t fix our fundamental incompatibility.”

## Chapter 6: Clear Waters

Tom attended Gatsby’s party and immediately announced: “I distrust you and will investigate your background.”

“Fair enough,” Gatsby replied. “I am indeed hiding things.”

Daisy pulled Gatsby aside. “These parties make me uncomfortable. You’re not the same person I loved five years ago.”

“You’re right,” Gatsby admitted. “But I’m unable to accept that.”

## Chapter 7: Direct Confrontation

In the hotel room, everyone laid their cards on the table—literally, they brought actual cards with their feelings written on them.

Tom’s card: “Gatsby is a criminal.”
Gatsby’s card: “Daisy loves me, not you.”
Daisy’s card: “I love security and comfort more than either of you.”
Nick’s card: “I’m uncomfortable and want to leave.”
Jordan’s card: “This is entertaining.”

“Well,” Tom said, “since Daisy has clarified her priorities, and I offer more security, I win.”

Gatsby nodded sadly. “Your logic is sound.”

## Chapter 8: Forward Motion

Driving home, Daisy didn’t swerve or panic. She drove straight and responsibly, arriving home without incident.

“That was anticlimactic,” Gatsby said.

Later, Wilson approached Gatsby at his pool.

“Did you have an affair with my wife?” Wilson asked directly.

“No, that was Tom Buchanan. He just told you it was me to deflect blame.”

“Oh. My mistake.” Wilson left to find Tom.

This is, in fact, what early acceleration of science looks like with jagged AI.

Writing abstracts takes a lot of time, many scientists are bad writers and clearly written abstracts are important. As long as the authors check over the results, this is a good use case for AI.

https://twitter.com/dkthomp/status/1951677835124330949

I wrote about some of the early impact of AI on science last year, including for writing. oneusefulthing.org/p/four-singula…

Kinda amazing: the mystery model "summit" with the prompt "create something I can paste into p5js that will startle me with its cleverness in creating something that invokes the control panel of a starship in the distant future" & "make it better"

2,351 lines of code. First time

https://twitter.com/emollick/status/1922749136996114771

Code: editor.p5js.org/emollick/sketc…

Play it: editor.p5js.org/emollick/full/…

Don't leave AI to the STEM folks.

They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI These are systems that respond to human writing and (often) techniques that apply to human psychology.

Everyone now has a machine that makes words, images, video, sound where the limit is often your own ability to imagine something new (or invoke old ideas others do not know).

X (and other social media sites) make our 1990s optimism about the Information Age seem silly.

Even with all of the world's information a click away (& a free AI that can help explain that information in a personalized way), half-mangled anecdotes with no source win every time. It really is not what most people who was working on building the early web in the late 1990s were expecting. Universal access to information was going to transform everything, creating widespread learning and bridging divides.

It really is shocking how much that didn't happen.

So, OpenAI Deep Research can connect directly to Dropbox, Sharepoint, etc.

Early experiments only, but it feels like what every "talk to our documents" RAG system has been aiming for, but with o3 smarts and easy use. I haven't done robust testing yet, but very impressive so far. I think it is going to be a shock to the market, since "talk to our documents" is one of the most popular implementations of AI in large organizations, and this version seems to work quite well and costs very little.

Very big impact: The final version of a randomized, controlled World Bank study finds using a GPT-4 tutor with teacher guidance in a six week after school progam in Nigeria had "more than twice the effect of some of the most effective interventions in education" at very low costs


Paper: documents.worldbank.org/en/publication…

Microsoft keeps launching Copilot tools that seem interesting but which I can't ever seem to locate. Can't find them in my institution's enterprise account, nor my personal account, nor the many Copilot apps or copilots to apps or Agents for copilots

Each has their own UIs. 🤷‍♂️ For a while in 2023, Microsoft, with its GPT-4-powered Bing, was the absolute leader in making LLMs accessible and easy to use.

As someone who has spent a lot of time thinking and building in AI education, and sees huge potential, I have been shown this headline a lot

I am sure Alpha School is doing interesting things, but there is no deployed AI tutor yet that drives up test scores like this implies. I am not doubting their test results, but I would want to learn more about the role AI is playing, and what they mean by AI tutor, before attributing their success to AI as opposed to the other dials they are turning.

AI tutoring is still being understood

https://x.com/emollick/status/1901487968105197968

I suspect that a lot of "AI training" in companies and schools has become obsolete in the last few months

As models get larger, the prompting tricks that used to be useful are no longer good; reasoners don't play well with Chain-of-Thought; hallucination rates have dropped, etc. I think caution is warranted when teaching prompting approaches for individual use or if training is trying to define clear lines about tasks where AI is bad/good. Those areas are changing very rapidly.

“GPT-4.5, Give me a secret history ala Borges. Tie together the steel at Scapa Flow, the return of Napoleon from exile, betamax versus VHS, and the fact that Kafka wanted his manuscripts burned. There should be deep meanings and connections”

“Make it better” a few times… It should have integrated the scuttling of the High Seas Fleet better but it knocked the Betamax thing out of the park

🚨Our Generative AI Lab at Wharton is releasing its first Prompt Engineering Report, empirically testing prompting approaches. This time we find:
1) Prompting “tricks” like saying “please” do not help consistently or predictably
2) How you measure against benchmarks matters a lot
Using social science methodologies for measuring prompting results helped give us some useful insights, I think. Here’s the report, the first of hopefully many to come. papers.ssrn.com/sol3/papers.cf…

The lack of benchmarks for writing, telling stories, persuasion, creativity, emotional intelligence, perceived empathy, and doing office work are...

(1) holding back AI advances, (2) hiding big differences between models & (3) obscuring how good these models are for real work If you want to influence the future, now is the time to release a really good benchmark.

https://x.com/jamescham/status/1894410097931198464

The significance of Grok 3, outside of X drama, is that it is the first full model release that we definitely know is at least an order of magnitude larger than GPT-4 class models in training compute, so it will help us understand whether 1st scaling law (pre-training) holds up. It is possible that Gemini 2.0 Pro is a RonnaFLOP* model, but we are only seeing the Pro version, not the full ultra.

* AI trained on 10^27 FLOPs of compute, an order of magnitude more than then GPT-4 level (I have been calling them Gen3 models because it is easier)

There is a lot of important stuff in this new paper by Anthropic that shows how people are actually using Claude.
1) The tasks that people are asking AI to do are some of the highest-value (& often intellectually challenging)
2) Adoption is uneven, but many fields already high

Paper: assets.anthropic.com/m/2e23255f1e84…