iDon

I'd agree with other posters - I'm seeing less hallucination; probably the model-makers are tuning to reduce that, because people make a big deal about it.

I see it as a small deal - it reminds us to check the responses against references. LLMs are a statistical construction, and everyone accepts without complaint that statistical models have a predicted false-positive and false-negative rate. I think of "hallucination" as a false positive; false negative is no answer when the model could have made a useful response. I think there's some relation between our creativity and the LLM capability we dismissively call hallucination.

The Z80 has a relative jump. Teenage me was loading a hand-assembled search function as an initialised array of integers in Basic, and relative jumps didn't need to be adjusted for the load location, which reduced the amount of post-load adjusting of addresses. It was +100x faster than Basic, just a straight scan.

> ... It starts by stepping into a shallow pool of golden light. > ... Together they act as both a choir and an audience

I think I'm not the target audience. I guess they are going to need to sign up a lot of people, to train on their scans + their medical outcomes. So the article is talking to people who will get enthused by it, which is more difficult after the question of 23AndMe data sale.

The risk is humans using AI to control / exploit / coerce / injure other humans. The risk of AIs being given enough agency to threaten humans comes after that - they'll only have the agency we give them (being "alive" or "conscious" is not the near-term risk).

The article lists diplomatic actions which might help to manage the risk, starting with "An agreement between America and China" - they all sound like impossible dreams. We had ~80 years of relative peace and prosperity in which to construct a framework of unity to face challenges like AI (and global warming, which until GPT I thought was the bigger risk); international unity is weaker than ever. In geopolitics and defence, capability of other nations is the concern rather than intention; the capability curve of LLMs is heading off our charts. We're backed into tight corners on nuclear proliferation, global warming, and I can see LLM-enabled conflicts (cyber warfare, infrastructure terrorism) pushing us over those other edges. Our democracies seem weakened, and I expect LLMs will empower those using social media to create conflict and control opinion. We're familiar with the cycle of inventing new technology which benefits people, then seeing how long before people invent ways to misuse it. There is a possibility here that LLMs could be used to solve the problems we are juggling, but I struggle to imagine that people won't misuse it even faster.

The article is a start on thinking and talking about managing our risks. The best outcome would be it is so well managed that, like the Y2K "bug", people say "after all that hoopla, nothing happened". I'm not seeing a smooth path to there.

I have drafted a photo gallery app in Rust / Tauri, using a JavaScript framework in the frontend. The backend can read directories and files directly, and because the backend and frontend are in a single process, the backend simply passes a file handle (path string possibly) to the frontend. In contrast Electron has to send the image file between processes. I started with Electron and I think that was the point I shifted to Rust / Tauri; seeing the images display immediately was a revelation. Rust / Tauri has the advantages of a desktop app, and I have the option to use the frontend as a web app also.

This Python binding (pytauri) is interesting too - I have colleagues with Python functionality they want to surface on the web, and this would give the possibility of running as a desktop app also - good for large datasets.

A search for : "character studies of old farmers" yielded good results. The results are drawings / engravings, which may reflect the balance of the collection, and perhaps this subject is more used in practice than in marketable oil paintings.

Since this is a semantic search, using a vector embedding, it will handle meanings better than a text search, which would handle names better.

The article asks what is the meaning of OpenAI's statement that "the UAE will become the first country in the world to enable ChatGPT nationwide."

My first guess would be that it will be a geo-fenced service, in particular UAE residents will have (subsidised) access to it and perhaps not to the global service, and it will have a system prompt designed and tuned in consultation with the UAE government.