Ethan Mollick

A gap for Claude is a lack of easy sharing. This especially true for artifacts. Non-technical people can create useful little tools, but don’t have the knowledge to easily share those with others or have them work sustainably outside of Claude (Devin solves this by using Netify)

Time to bring back personal websites for hosting tiny webapps. Geocities anyone?

Professor @Wharton studying AI, innovation & startups. Democratizing education using tech
Book: https://t.co/CSmipbJ2jV
Substack: https://t.co/UIBhxu4bgq

Entity Service top level domain, every item that is in Entity Service should be in this domain

Entities [Entity Service]

Services

Brand

Google 

Products created by Brands.  Examples: Ford Explorer, Apple iPhone.

Product

Google brand conversation

A grouping of interests and hobbies entities, like Novelty Food or Destinations

Interests and Hobbies Category

Horoscope

Interests, opinions, and behaviors of individuals, groups, or cultures; like Speciality Cooking or Theme Parks

Interests and Hobbies

Gemini

Unified Twitter Taxonomy

Linear growth in the amount of GPT in GPT.

And yet Claude and Google seem to have slowed down, increasing their amount of Gemini and Claude by only .5 per release. 

Exponential growth requires Llama 9 to be released next 😉 https://t.co/MgXEWEM8Cx

GPT series still hasn't shown any signs of saturation 😲 https://t.co/TIBRYVWCEk

Aran Komatsuzaki

Knowing Twitter I will clarify that this is a joke.

When I gave Claude a picture and told it:

"ENHANCE (play along, use artifacts)"
"More drama ENHANCE:
"YES! ENHANCE!"
"ENHANCE HAHAHAHA! YES!"

Here are the four interfaces it generated. https://t.co/4x2dLxAOar

The text is also pretty fun. https://t.co/HolapUt18z

High value things for non-coders that LLMs could do today but frontier labs are not doing (maybe because they are all coders):
1) Handle excel sheets as they are in real life: formulas, weird layouts
2) Write directly on PDFs or docs (fill out forms!)
3) Better finance training

Stop maximizing MMLU and coding challenges and pick some metrics that matter to most people to maximize!

Categories within Brand Verticals that narrow down the scope of Brands

Business Taxonomy

Brands, companies, advertisers and every non-person handle with the profit intent related to softwares, apps, communication equipments, hardwares

Technology Business

ChatGPT

Technology

Artificial intelligence

for individual and types of technology, e.g., food technology, 3D printing

You can get GPT-4 to do these things, but it is annoying. Make it not annoying!

"So much depends on the red wheelbarrow, glazed with rain. Claude 3.5 please create a simulation to show me why"

I appreciate that the farmer's health declines with no wheelbarrow, but improves fastest when using a red wheelbarrow, glazed with rain because of aesthetics. https://t.co/3pobFAY0CD

Please create a simulation of canto IV in The Wasteland.

I really appreciate the literalness with which it handled "And the profit and loss." by including an actual profit and loss meter. https://t.co/JqcF0px8jd

Just going to keep getting more obscure from here.

Simulate Archibald MacLeish's Ars Poetica. https://t.co/mHfvr7H8vE

Animate "Little Gidding" from The Four Quartets

I realize that I have lost everyone at this point, but it is pretty neat:

In the words of Claude: A small circle in the center represents "the still point of the turning world," and the pink circle is the rose, the green the yew. https://t.co/kbXhBdYDfN

Brands, companies, advertisers and every non-person handle with the profit intent related to offline and online games such as gaming consoles, tabletop games, video game publishers

Gaming Business

Gaming

Adventure games

"The Lady of Shalott" as an adventure game. https://t.co/tgr2MPSIvo

I have reached the level of intersection between AI and classic poetry where my audience drops to zero.

Politics

News

There are some papers finding LLMs are more persuasive than most humans, with unknown consequences for politics. So this paper was some relief, because it suggested super-persuasion didn't scale...

...but I am not convinced their prompting is going to maximize persuasiveness https://t.co/PdcwChxp13

"AI can't do something" papers are a hard genre, and I do think they do a good job given constraints, but I would love to see some testing of their prompts to maximize persuasiveness before they put them in the field.

Also, most of the hyperpersuasion work comes not from static messages, but from interactions with the AI. I am not sure writing ads is the same thing as debating or discussing with the LLM.

Video games

Simulation games

When you ask Claude 3.5 to make a snake game, it has lots of examples in its training data, but it still does a good job when you ask it to create something novel and complicated: "make an interactive simulation that explains tertius gaudens and robust action" "make it better" https://t.co/W39B7PKUM0

"i need you to create a simulation where I am a lighthouse keeper in a cosmic lighthouse and I have two beams one that attracts space leviathans and the other repels them and there is also gravity to take into account. make it fun and make the graphics good and also add lore and make it have a goal" (followed by a little debugging)

"i need you to create a simulation where I am a lighthouse keeper in a cosmic lighthouse and I have two beams one that attracts space leviathans and the other repels them and there is also gravity to take into account. make it fun and make the graphics good and also add lore and… https://t.co/9YJuVXzhRg https://t.co/3uSCGLRCIN

What can be so hard to grasp with AI systems is their inherent randomness.

I gave Claude the prompt "create a webgl artefact that is a social networking simulation, is interactive and explains brokerage" five times

Once the sim didn't work, twice it was fine, twice it was great https://t.co/UKavWdWDLc

In some cases it insists it can't run webgl

Ideas come from the edges, not the center

Latent Expertise: Everyone is in R&D

Experts in a field are going to be the best users of AI in that field.

I put a few examples in the post but I really like using the 10+ page prompt by Glowforge co-founder Dan Shapiro, an expert in company culture, for making job descriptions inspiring. https://t.co/FqslPLu8T5 https://t.co/eKaOaGdvai

In many cases, prompting efficiently and knowing fancy LLM tricks is less important to using AI than knowing how to do something and being able to communicate what you know.

We are at an interesting point where LLMs are doing stuff that is genuinely astonishing for a next token prediction system, but the stuff it is doing seems reasonable to many: “of course AI can build a teaching simulation and summarize documents and hey look at the crude drawing” https://t.co/K9APHoSZDR

Drawing reference

"Claude 3.5, here is a 78 page PDF. Create an infographic describing its major findings." (accurate, though the implications are its own)

"Claude 3.5, create an interactive app demonstrating the central limit theorem"

"Claude, re-create this painting as an SVG as best you can" https://t.co/fXuujzXrIR

@TommyFalkowski @emollick When Stable Diffusion came out a couple of years ago I remember seeing comments from people who had shown it to unimpressed non-techie friends who said "I thought that was what Photoshop did?"

Creator @datasetteproj, co-creator Django. PSF board. @nichemuseums. Hangs out with @natbat + @cleopaws. He/Him. Mastodon: https://t.co/t0MrmnJW0K

@emollick's Latest Twitter Threads

Read threads by Ethan Mollick and get notified about future ones.