Paul Tiplady

CTO @ Qwil

[Paul Tiplady @EndpointZero]

@dev__cycle @sun_hanchi @AnthropicAI Right. They fronted “up to $100m” for this. Arguably they could get better goodwill by giving one big fix to every OSS project, rather than focusing exclusively on the big ones. But given limited time the big ones are the place to start.

[Paul Tiplady @EndpointZero]

3. More speculatively, beyond traditional Merge Trains, the Agentic version is to dispatch a fast & cheap agent to fix merge/bug failures dynamically before aborting and returning the PR to the author; a lot of minor fixes don’t need lots of context or big models to repair.

[Paul Tiplady @EndpointZero]

https://t.co/SMPL0WIX0n

Loving this post from @Steve_Yegge , a lot of interesting material. I agree with a lot of the agent swarm observations. One item I’m gonna debate is giving up on CI/CD. Might make sense for games; not sure it will generalize to enterprise software.

[Paul Tiplady @EndpointZero]

2. Not viable for games, but beyond blue/green deploys you can get more parallelism with Rainbow deploys; each train rolls to production concurrently. Probably too confusing for humans to keep track of but agents can do a better job here. Requires more discipline with migrations.

[Paul Tiplady @EndpointZero]

1. merge trains let you batch a lot of commits & speculatively advance your CI HEAD state. I suspect this can solve a lot of the throughput issues.

[Paul Tiplady @EndpointZero]

@antirez I don’t think it’s a coincidence; rumination seems to be a generic algorithm for search in concept-space. When you include temperature (as human rumination also does) it’s a kind of genetic algorithm; you randomly connect concepts and can produce new, fitter combinations.

[Paul Tiplady @EndpointZero]

@BruzWJ @aslater My friend, let me tell you about this thing called a “Trade Mark”. You register your logo with USPTO and then they are protected. So yes, according to federal law, that is a logo.

[Paul Tiplady @EndpointZero]

@pitdesi Seems likely to me that vertical integration is the way forward. Why give API access to your competitors so you can get distilled? Sell the end product not tokens. The coordination costs of running a conglomerate are only going down. Post-Mythos, this path is even more appealing.

[Paul Tiplady @EndpointZero]

@MakeUSAVAT @quid_pro_quore @tszzl You know that “your best move given your objectives” != “my preferred outcome”, right?

[Paul Tiplady @EndpointZero]

@EpochAIResearch Seems likely that running in Claude Code (with explicit skill authoring instructions) would be far superior than a bare ReACT loop. Most of Claude’s training for ICL is specifically using skills and tools.

[Paul Tiplady @EndpointZero]

@TomLoweCinema @AndrewCurran_ The EUV tech stack was a merger of ASML and major US contributions. So USG has export waiver veto rights on it. https://t.co/FbveB6kWC8

[Paul Tiplady @EndpointZero]

@slatestarcodex @drethelin There’s a bit of a chicken-and-egg here; the iatrogenic risk of imaging is downstream of broken diagnostic incentives and poor quality (high FPR) imaging. Presumably at some quality bar this flips and the over-treatment harms no longer dominate?

[Paul Tiplady @EndpointZero]

@antirez I think it should be fine if we conceive of HTML as the presentation layer, the leaf nodes in the knowledge graph; a denser representation is still needed for the underlying ontology.

[Paul Tiplady @EndpointZero]

@MendozaVictor50 @cremieuxrecueil @grok correct these numbers for Socioeconomic Status and other population bias confounders.

[Paul Tiplady @EndpointZero]

@karpathy I like this pattern. Two things I added recently - 1) sync Obsidian to GitHub. Then Claude Code iOS can natively load & work on your KB repo. 2) GitHub action to periodically refresh articles with Claude summarization: https://t.co/u9C4iIMugc

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

[Paul Tiplady @EndpointZero]

@karpathy I feel this story and many of Greg Egan’s works share a beautifully (cynically: naively) optimistic vision of inter-species contact; all intelligences are curious and friendly by default, collaboration is easy (especially by engineer/scientist types). Counter argument: biology.

[Paul Tiplady @EndpointZero]

@felixrieseberg Would be great if I could transfer chats between modes, lots of times what starts as a quick discussion escalates to “build this durable artifact” and I don’t see a way to Dispatch from an existing chat. I often hit the same problem with Chat -> Code.

[Paul Tiplady @EndpointZero]

@PeterJ_Walker @pitdesi Is a vintage analysis the right lens here? Eg don’t expect to see a few-year old trend on Series D companies. The current averaging might be masking stronger effects in the most recent years.

[Paul Tiplady @EndpointZero]

@karpathy It feels like the original vision of git itself may have something to offer; swapping patches and branches between each contributor’s repo, with a Linus merging every winning concept back upstream. I wonder if we even need GitHub here or if an email list would work for sharing.