Recent Posts

v cool demo examples

See for example comments on benchmarking

Discussion better than article etc

Fable prompts - way more ambitious than my current side projects (with Codex)!

Related:

Example of the pitfalls of vibecoded handbooks

Extremely good for soft skills & business impact (see especially: "Defining DevRel at Google" at 1:29:20)

See also: First 90 Days

Recommended by a generative model researcher I know (see also: Karpathy's GPT2 youtube series)

see also: cattle, not pets

(based on 2024 data)

Lots of debates on utility, but prob will have some demand from consumers but tbd on medical industry

RIP

tl;dr: good idea for those 50+ to get the shingles vaccine; also seems to have ~20% reduction in dementia diagnosis

Equally / more so applies with AI coding

It's a hard problem! In my side project I also use a consistent, game-local seeded RNG so runs/simulations/etc are all fully deterministic & repeatable. But leads to challenges like the above.

"Champions of a European AI model should ask themselves if a European effort would be more effective than Meta, which this year will spend more on chips ($125 billion) than Germany spends on defense ($114 billion) and offer salaries of over $100 million to attract the best researchers, and is still failing to catch up. Elon Musk tried and failed to build a good AI model."

Sarah Paine, more historical background here (vs just modern history)

Codex $100 is plenty for even dedicated hobbyist coding IME, but if you're doing a startup I would get double $200 subs (OpenAI + Anthropic) and go from there. Paying API rates for me at work is thousands per month spend.

OpenRouter trying to standardize the common practice of blending expensive & cheap models. Also interesting is the "Preventing Models From Cheating" section, b/c it is a bit of tooling/tech to be able to run evals on models that can access eval-related data/solutions through general search (web, code search, or internal company search tools)

Roughly endorse the first couple sections; a prior manager I had also recommended having 20% slack [at work] - which can be used for reading, thinking, etc; and when it’s crunch time can ofc allocate that as needed Lately I’ve been using my “20% time” at work to be putting out one side project or another, typically done by a Claude or two in the background of my main work. Not sure on the implications of that for having real slack time to think/process/etc in the background

Some of this philosophy is why my personal website is a simple static site (but ofc I can embed JS plugins so not entirely)

p classic prompt injection

Jailbreaking LLM auto analysis by adding nuclear & bio weapons text, so it triggers LLM safety refusals

Interesting interview with a well known name about the hedge fund industry, starting with an introductory 201-level explanation of what a quant is, what hedge funds do, etc. Recommended by a Citadel analyst

Quality is the real challenge, not orchestration technical capability IMO

People are using this to turn 2d AI art to 3D to then use in game dev / animations. Both directly (3D game etc) and indirectly (3D model, rotated and posed in different ways to make a consistent 2D spritesheet, which is otherwise hard / requires LoRA etc)

His next finished series

By Knausgaard

Shreya is always excellent

Pretty classic, it's why I got my new PC early last year (also pre-tariffs) and new phones earlier this year In general suggest getting your long-lived consumer hardware now, b/c it will prob take 1-2y (or more) before prices drop meaningfully. And rn many manufacturers are still semi-subsidizing/not pricing in cost increases just yet due to manufacturing & stocking timelines (e.g. recent Switch price increases)

comments better than article

Related: residential proxy abuse makes scraper traffic harder to block.

Retraglutide is achieving bariatric-surgery-scale results

"Apparently, in the era of AI, there are a lot of delusional old people out there! They can't tell fact from fiction, but in many cases have unwarranted confidence in their ability to do so. This seems like a very bad combination, and indicates that the elderly will be falling for a lot of scams over the next few decades, in the same way they fall for non-AI internet and telemarketing scams."

The specific exhaustion of being a mother.

tl;dr: Restaurant patrons in a recent study tipped their servers up to about 1 percentage point less on Saturdays and Sundays.

Some inside baseball "Let me paint two scenarios. In one, people are fine with AI-generated slop despite their protests, the platforms shrug, and feeds become a slop war. In the otherwhich I think is more likely people hit a threshold where the internet enshittifies to the point of being unusable and they retreat into private spaces with trusted voices. If that happens, parasocial relationships with creators will hit warp speed. People will seek out the creators, influencers, and voices they really want to hear from by going to them directly. And their real conversations move further underground into group chats, WhatsApp groups, Instagram DMs, Snapchat, Discord servers." I also agree the second scenario is more likely - see also Imas's articles on 'what is scarce', relational economy, etc

Related: Amazon delivered more parcels in 2025 than the Big 3 domestically in the US.

Interesting wrt incremental derisking (pop ups, cooking show, then restaurant). Also Nishaan runs off social media (TikTok, IG) - tourists visiting NYC (including esp muslims) go specially to it - can see that mentioned frequently on reviews

related to the main discussion here

"With large gains in life expectancy, the population share of disability due to cognitive decline and dementia has substantially increased. Many older adults in the United States leave the workforce well before age 65. Correlational evidence suggests that leaving the workforce before retirement age could accelerate the pace of cognitive decline. We offer causal evidence, using HRS data for the United States, exploiting plausibly exogenous shifts in labor demand in local labor markets as a Bartik instrument for employment variation across these markets. We find substantial declines over time in cognitive scores stemming from negative labor demand shocks. These findings are concentrated among men aged 51 to 64, whose employment decisions and outcomes may be more sensitive to local labor market conditions than are these decisions or outcomes for women or for older men. Our evidence extends past work focusing narrowly on the retirement age window and provides further support to the notion that working to older ages may delay age-related cognitive decline."

Big step for OpenAI to compete for corporate contracts (wrt regulations, easier procurement, etc)

Prediction markets in everything

Lot of people are working on how to use LLMs effectivey in writing prose (as opposed to code) - see Gwern, romance mass production novelists, this article, etc. Mostly around the harness layer, tho they (except for Gwern) might not phrase it that way Prob low hanging fruit in cross applying learnings & techniques from coding harnesses to prose

[11:50 PM, 4/20/2026] Veer

Good advice

original paper (above article seems LLM generated)

"Quince’s Best-Kept Secret Isn’t Its Cheap Cashmere and Couches"

also this; about halfway through and it’s interesting wrt what are the specific bottlenecks for scaling AI Stratechery also cites him (Dylan Patel, SemiAnalysis): Also aligns with this discussion

"5. A big win from using AI agents is tackling stuff that you wouldn’t have before. A senior engineer at 37signals ran a “P1 optimization” project to improve the fastest 1% of requests. They optimized the P1 from being at 4 milliseconds, to under half a millisecond. This is the sort of work that wouldn’t have been considered previously!” 100%, I usually have an “eng excellence” task going in the background while I focus on feature work. Lets me continuously try out new ideas, deliver improvements, clean up code, etc

Inference has ties to the Harness engineering as well - can have clever tricks like fine-tune the LLM to plan tool calls in parallel, which ofc relies on your harness supporting that capability (that said I’m a little tired of all the Foo Engineering catchphrases, we can just call it platform & infra eng with a focus on LLMs lol)

“GLM-5.1: Towards Long-Horizon Tasks”

We’ll get there tho, you can see e.g. Anthropic is working on the building blocks with subagents + coordination mechanisms, plan mode, Claude Code for the Web (containerized execution) - I would expect within 3 - 6m at most we’ll see full workflow / team orchestration as a product release from them Ditto for Codex MVP version might not be there tho even with strong eng teams, might need another model bump or two, or more “harness engineering” and thoughtful human in the loop. Maybe end of this year / early next Tough to really say given exponentials Here you go

see also: today’s Stratechery

also this; about halfway through and it’s interesting wrt what are the specific bottlenecks for scaling AI Stratechery also cites him (Dylan Patel, SemiAnalysis)

“Amazon in talks to buy $9bn satellite group Globalstar in bid to rival Elon Musk’s Starlink”

“OpenAI’s ChatGPT App Store Took Aim at Apple, But Results Lag So Far”

Mix of marketing hype and interesting claims; note that a GPT model version announcement is rumored later this week

“Employers are using your personal data to figure out the lowest salary you’ll accept”

Interestingly, it explains why my parents like Nescafe Gold more & why it costs more than the average instant coffee - it's made with freeze drying process rather than the cheaper conventional spray drying (around 2x the cost)

Good links

He used oh-my-codex for orchestration

“WNBA Players Had an Ace Up Their Sleeve in Pay Negotiations: A Nobel Laureate”

“China cracks down on ‘bone ash’ burials in empty apartments”

example usage

"What major works of literature were written after age of 85 75 65!"

Good talk!

Ted Chiang's "Catching Crumbs from the Table," a 1 page sci fi short story on humans learning from superintelligent metahumans

Key glue work for better agentic eng (especially greenfield work)

Experimental etc but lot of products out there doing AI design

Comments better than article