i got good feedback from the format last time so i'm going to take it up a notch.
in the text below anything written by ai is in plain text where as my commentary is lowercase, italic, and in a block-quote, with blemishes not edited out. let’s go!
🦾 How I used AI this week
i got a parking ticket. so naturally i took a photo of the front and back and gave it to o4-mini-high with the simple instruction "how can i fight this and win."
the model came back saying that the flashing red light on the meter that the ticket talked about could indicate a "meter malfunction" and i should request maintenance logs from the city to prove that the meter was actuality functioning.
the model said "Municipal Code § 3.16.250 states the city must prove the meter indicates time limit is over, but a broken meter can’t validly do so". the model seems to think that if the city can't prove the meter was working the ticket will get thrown out.
in this case i should have gotten the ticket (i forgot to feed the meter) so i’m just going to pay it. but it was a good reminder on how powerful these systems can be at providing at least some direction in an ambiguous situation.
i've helped people setup very low cost rag systems to do litigation support and the results were pretty amazing. i think that harvey or do not pay are just the tip of the iceberg.
there is a huge tam with people that need to do things like fight parking tickets, lawyers, or other civil (and even criminal) legal issues. previously these things would be taken on by pro-bono attorneys but imagine the impact the pro-bono firms will have with this technology.
🤖 AI
Accents in Latent Spaces: How AI Hears Accent Strength in English
tldr here is that the author was able to represent a skill gap in latent space and build a system to close the gap. they then took the persons voice and generated audio training recordings of the person without an accent.
the learner used the recordings as training and the results were pretty amazing (visit the link and listen).
imagine the impact that doing this for all skills will have on education. its something i’ve been thinking about and hacking on for years. the tooling is finally getting good enough to make this a reality.
Sakana AI: The Darwin Gödel Machine (DGM) represents a significant advancement in AI research, allowing for self-improvement through autonomous code modification. This approach, inspired by principles of evolutionary biology, enables the DGM to create and evaluate various coding agents, leading to substantial enhancements in performance on programming tasks, particularly in benchmarks like SWE-bench and Polyglot.
agents that write themselves in an evolutionary cycle...sounds cool…but very expensive.
it took billions of years of the sun sending energy to the planet before i could write the words you are reading right now. so i guess bio evolution is expensive as well.
An alleged leaked version of Claude's system prompt:
its kind of crazy to me how long this prompt is (11k lines). it would take me an hour at least to read and digest the whole prompt.
the claude models can read and understand in a fraction of a second. llms really are a different type of intelligence. there are also some good ideas here on how to structure prompts. <reminder>llms love xml tags.</reminder>
UnitedHealth Has 1,000 AI Applications in Production: UnitedHealth Group is significantly expanding its use of artificial intelligence, with 1,000 AI applications currently in production across its insurance, health delivery, and pharmacy divisions. AI technologies are being utilized for various functions, including transcribing clinician conversations, summarizing data, processing claims, and managing chatbots, with half of the applications using generative AI.
regression could be considered "ai"…. so "ai" is such a broad term…nonetheless we should start to see more companies who have an interest in reducing costs adopt frontier techniques which is exciting.
AI in the enterprise: OpenAI outlines key lessons for successfully adopting AI in enterprises, highlighting a shift towards an experimental and iterative approach.
one of the biggest struggles with ai adoption for companies is in being iterative.
not all projects bear fruit so you need to be willing to try and fail at a lot of things before figuring out what works. because of this people leading ai initiatives will need a lot of political capital within their organizations.
but for those who stick it out the roi is definitely there.
FLUX.1 Kontext: FLUX.1 Kontext is an innovative image generation model that expands beyond simple text-to-image capabilities by allowing users to create and modify existing images based on text instructions.
the image above for this email was done in this model.
💡Steal this idea
right now when you are doing an integration between systems coding agents can read api documentation and build the necessary connections.
this works pretty well if the interactions are simple.
a better approach would be for the company exposing its api to wrap the api docs in an endpoint that is an agent.
the agent would read through the docs then build / test / store it own implementation examples. then if you wanted to integrate with a system you could just have your agent phone the other systems agent and describe the problem its trying to solve.
this would allow the remote agent and you coding agent to collaborate on a solution (maybe even in a shared workspace). the remote agent could "learn" from this interaction and improve its set of implementation examples for future integrations.
🔗 Cool *hit
Onlook is cursor for designers: Onlook is an open-source design tool that allows seamless integration with any React-based website or web app, enabling live edits directly in the browser.
we’re going to see more and more of the “cursor for x” products pop up.
claude’s sequential tool calls is pretty powerful. one thing i’ve bee thinking about is what tools are needed for people who are vibe coding / sandcastle building.
there are lots of “dev tools” for vibe coders that just don’t exist yet (testing, security, deployment, etc).
you can also look at all of the other functions in an org that are not coding…what happens when you can do “vibe analysis” or “vibe customer service”… lots of stuff to build. not enough time.
in a world where high fidelity visuals can be generated, low fidelity visuals are still really cool. imagine that.
Clippy Desktop Assistant: Clippy is a desktop application that allows users to run large language models locally with a nostalgic 1990s user interface reminiscent of Microsoft design. Clippy supports custom models and settings and operates offline, making it free and local.
microsoft should have made this.
How to 3D Print Your Personalized AI Action Figure: The latest trend on social media involves creating personalized action figures using AI, especially through ChatGPT, which has become widely popular. Freelance writer Amanda C. Kooser details her amusing experience generating her own action figure, complete with a rock star outfit, and shares tips on how others can do the same.
On The Death of Daydreaming: In a thought-provoking introduction to Christine Rosen's new book, Jon Haidt highlights the detrimental effects of smartphone dependency on both Gen Z and older generations. She argues that the loss of interstitial time—those small gaps in our day previously filled with reflection or idle thought—leads to significant declines in social engagement and increases in anxiety.
there is something directionally correct here.
i have by most creative thoughts when: riding a bike, working out, in the shower, or walking. these are all times i’m unplugged.
i need more unstructured time.
🔈What I'm listing to
don’t judge plz
📚 What I'm reading
ὁ κόσμος ἀλλοίωσις, ὁ βίος ὑπόληψις.
"The universe is change; our life is what our thoughts make it." -Marcus Aurelius