edwin

There’s something quietly impressive about getting modern AI ideas to run on old hardware (like OP's project or running LLM inference on Windows 3.1 machines). It’s easy to think all the progress is just bigger GPUs and more compute, but moments like that remind you how much of it is just more clever math and algorithms squeezing signal out of limited resources. Feels closer to the spirit of early computing than the current “throw hardware at it” narrative.

I’m one of the authors.

Our interest is straightforward: If coding agents scaffold a growing share of new apps, their defaults will become prevalent. Not necessarily because they’re objectively best, but because the choices are frictionless. The model becomes the gatekeeper of early architectural decisions that used to happen in a meeting room.

So we measured what those defaults actually are.

We ran structured app-building prompts through Claude Code, captured the generated repos, and extracted stack choices: auth, UI framework, database, deployment assumptions, package management, etc.

A few observations that stood out: - Deployment assumptions are shifting toward Vercel and Railway, and away from AWS/GCP-first patterns. - Defaults evolve as models update. For example, Opus 4.6 recommends Drizzle more frequently than Prisma.

Prompts, raw outputs, and parsing logic are here: https://github.com/amplifying-ai/claude-code-picks

This is a snapshot in time. An interesting question is how quickly these defaults drift.

Unlike classic search, which got worse over time due to SEO gamings, AI search might actually improve with scale. If LLMs are trained on real internet discussions (Reddit, forums, reviews), and your product consistently gets called out as bad, the model will eventually reflect that. The pressure shifts from optimizing content to improving the product itself.

We looked at the same data that Rand Fishkin used and definitely came to a different conclusion.

A few take-aways from a study we ran (~800 consumer queries, repeated over a few days):

* AI answers shift a lot. In classic search a page-1 spot can linger for weeks; in our runs, the AI result set often changed overnight.

* Google’s new “AI Mode” and ChatGPT gave the same top recommendation only ~47 % of the time on identical queries.

* ChatGPT isn’t even consistent with itself. Results differ sharply depending on whether it falls back to live retrieval or sticks to its training data.

* When it does retrieve, ChatGPT leans heavily on publications it has relationships with (NYPost and People.com for product recs) instead of sites like rtings.com

Writeup: https://amplifying.ai/blog/why-ai-product-recommendations-ke...

Data: https://amplifying.ai/research/consumer-products

Here you go: https://preview.promptjoy.com/apis/m7oCyL

That's not the current OpenAI recipe. Their expectation is that your custom data will be retrieved via a function/plugin and then be subsequently processed by a chat model.

Only the older completion models (davinci, curie, babbage, ada) are avaialble for fine-tuning.

With LLMs, the inputs are highly variable so exact match caching is generally less useful. Semantic caching groups similar inputs and returns relevant results accordingly. So {"dish":"spaghetti bolognese"} and {"dish":"spaghetti with meat sauce"} could return the same cached result.

New techniques like semantic caching will help. This is the modern era's version of building a performant social graph.

BTW: Here's a more performant version (fewer tokens) https://preview.promptjoy.com/apis/jNqCA2 that uses a smaller example but will still generate pretty good results.

Thanks. We find few-shot learning to be more effective overall. So we are generating additional examples from the provided example.

For those who want to test out the LLM as API idea, we are building a turnkey prompt to API product. Here's Simon's recipe maker deployed in a minute: https://preview.promptjoy.com/apis/1AgCy9 . Public preview to make and test your own API: https://preview.promptjoy.com

This is an exciting emerging space with folks (including multiple YC companies) tackling the problem from different directions. I helped compile a list here: https://medium.com/@vikati/the-rise-of-data-monitoring-7221e...

http://news.mit.edu/2016/how-cancer-cells-fuel-their-growth-... - this more recent study theorizes that amino acids, rather than glucose, supply most building blocks for tumor cells

I believe USDC has a blacklist of "forbidden addresses" - this is in contrast with some other compliant ERC20s that have whitelists instead. To redeem USD from USDC (meaning withdrawing said USDC from circulation), the withdrawer goes through the same AML/KYC process with Coinbase or Circle before receiving USD in a bank account.

DappRadar is a useful but ultimately hand-curated list. For a more comprehensive look, https://blockspur.com/ethereum_contracts/transactions ranks every Ethereum smart contract by transactions, "users," and "revenues." The top 3 contracts with the most number of users in June had 370K+, 210K+, and 88K+ "users" respectively. (Disclaimer: maker)

Greg Galanos (founder of http://en.wikipedia.org/wiki/Metrowerks) and his associate Mark B Johnson (Apple alum).

Last year, my 8-year-old kids and their classmates designed and hand-built (eventually with dowels instead of screws) their own chairs at their amazing SF school. See http://www.sfbrightworks.org/2013/11/hawk-chairs/

Here's a quote about the lack of undo: "The consequences of careless work were immediate and unavoidable, and the amount of focus, resilience and perseverance required by these chairs was immense."