Lak Lakshmanan

Screen door series by Jamie Wyeth of Andy Warhol by

Crib made of prison-like fencing by Hayden.

Also Hayden. Plywood ribcages on a new york subway rod.

Hayden. Cast iron pots and African masks

The Frye museum on First Hill has a couple of very good Americana exhibitions going on now. Strongly recommend. The alt-text in the images explains what they are
#seattle

Cameras used to be considered by painters as shortcuts and traitorous to the art ... They opened up new uses

But the big tech firms that have been responsible for the bulk of recent layoffs are highly profitable. Not sure how much the tax treatment was a contributing factor. (as opposed to making space for AI data center spend, for example)

How to Choose the Architecture for Your GenAI Application A framework to select the simplest, fastest, cheapest architecture that will balance LLMs’ creativity and risk

Putting it behind a paywall is a choice that the author made. I write on medium and my articles are free and can even be read in and incognito window. Try this: medium.com/data-science...

Thanks to all the AI engineers who shared what was working and what wasn't. And to Glen Yu , David Cardozo, Andrew Stein, Sarah Grey for their prompt and helpful reviews.

GitHub - lakshmanok/generative-ai-design-patterns: A catalog of design patterns when building generative AI applications A catalog of design patterns when building generative AI applications - lakshmanok/generative-ai-design-patterns

The patterns' code and brief descriptions are on GitHub:

github.com/lakshmanok/g...

Looking forward to hearing from you! What did you find helpful? Where can we do better? This is early access, so you still have a chance to make it better for future readers.

Generative AI Design Patterns Generative AI enables powerful new capabilities, but they come with some serious limitations that you'll have to tackle to ship a reliable application or agent. Luckily, experts in the field … - Sele...

The first five chapters of our 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐃𝐞𝐬𝐢𝐠𝐧 𝐏𝐚𝐭𝐭𝐞𝐫𝐧𝐬 𝐛𝐨𝐨𝐤 are up on the O'Reilly website. Labor of love from me and Hannes Hapke, but also a great learning opportunity as we got to explore things we didn't personally use.

www.oreilly.com/library/view...

Looks like an O'Reilly book cover

Cute story. The guy who's stuck tells the story tongue in cheek. Too many haters in the comments.

Wildflowers now blooming in western Washington trails! L to R: Coltsfoot, Trillium, Streaming Violet, Salmonberry

The graph shows change from the same 30 days a year earlier. So it's YoY

I was not able to replicate this, at least on Gemini or Claude
g.co/gemini/share...

claude.ai/share/00cb35...

Optimizing to the Eval GenAI Design Pattern #5

In this GenAI design pattern, I demonstrate how to use preference tuning to produce optimally performing content.

medium.com/data-science...

Very much what Kahneman called "system 1" and "system 2" in "Thinking Fast and Slow". System 1 is intuitive, system 2 is deliberate, and there are lots of ways that system 1 gets tripped up

Hold on your outrage. They just show the map per the official edicts. Indian and Chinese users do not even see the same border.

There are technical challenges to making self-play work for things other than games. This is legitimately advancing the state of the art

bsky.app/profile/lakl...

They open sourced their entire training regimen. Deepseek trained their model from scratch using the same technique that AlphaGo used ("self-play") and that requires both less compute and less labeled data. They are legit.

DeepSeek is now ranked just below the leading, recently deployed versions of Gemini and ChatGPT on LMArena. So how? Secondly, distillation requires access to so many input-output pairs that OpenAI/Google/Anthropic would immediately recognize and shut then down. (3/3)

They focused on easily verifiable problems. Which also makes sense given that the DeepSeek team is China's version of Jane Street. The people claiming the Chinese copied are thinking that they must have distilled their model from GPT-4 or Gemini but there are two problems with this theory: (2/3)

Additional context is that much of the cost of training foundational models like GPT-4, Claude Opus, and Gemini Pro is for the compute hardware (GPus/TPUs) and human raters/annotators. So DeepSeek managed to cut both these costs by getting pure Reinforcement Learning to work. To do that, (1/3)

They open sourced their entire training regimen. Deepseek trained their model from scratch using the same technique that AlphaGo used ("self-play") and that requires both less compute and less labeled data. They are legit.

Hair ice that we saw on our weekend hike near Seattle.

It is a rare type of ice formation caused by a specific fungus releasinv moisture from a specific kind of rotting wood when the weather conditions are just right: www.discoverwildlife.com/plant-facts/...

First hike of the year was in our backyard. This is Far Country Falls, a seasonal stream on cougar mountain. #seattle #wta

"Supercharged AI Intern" for your Product Management Team with NotebookLM Product Management is a demanding role. You are constantly juggling market research, stakeholder management, product roadmaps, user feedback, constantly changing engineering schedule, technology disru...

LLM-powered workspaces are the next new office productivity tool. This excellent article clarifies how to organize your sources, transcripts, thoughts, etc. and extract insights into templated formats (like PRDs) using one such tool. The underlying ideas are applicable to any white-collar work.

Why you’re not on holiday in India right now A fabulous destination for foreign tourists does little to lure them

Things India could do to make itself friendlier to tourists:
(1) Enable easy access to UPI to foreigners or via credit cards
(2) Allow e-SIM purchases based on passport, i.e. without aadhar/PAN card
(3) Make websites accept foreign credit cards and send OTPs to foreign ph numbers to enable planning

This is apple defending its 30% take rate even when there is no direct payment being made.

Mobile is 64% of all web traffic, and iOS is about 30% of that. If you pencil it out, iOS revenue to Google is about 25% of $240B or 60B. So, the $20B works out to about 30% of Google ad revenue on iOS. Exactly what apple charges for in-app payments. Google is paying the apple tax on its ad revenue

Evaluation-Driven Development for agentic applications using PydanticAI An open-source, model-agnostic agentic framework that supports dependency injection

Ideally, you can evaluate agentic applications even as you are developing them, instead of evaluation being an afterthought. For this to work, though, you need to be able to mock both internal and external dependencies of the agent you are developing. Dependency injection in PydanticAI unlocks this.

YouTube video by Raghuraman.V Rare video of G.N.Balasubramaniam singing in a concert//Yochana kamala lochana//#gnb

Is this AI-generated? People have long wanted to see GNB sing and Palani play, but they were before the movie era. And then this clip shows up. What do you think?
youtu.be/S6bx-zWnuwI?...

As you well know, execution (not the idea) is everything in tech. Clippy was built using rules and could barely check spelling, let alone grammar. This is like people saying Leonardo was ahead of his time because he drew flying machines. Yes, but the Wright Brothers actually got a plane in the air.