Chat GPT
-
I've been trying some of the Retrieval Augmented Generation (RAG) models. These prioritize specific documents you specify. Quite impressive when applied to scientific papers.
They can take data/conclusions from tables already, if they can interpret figures (graphs) then they will be very useful.
-
-
Fed some SQL in yesterday asking it to put the commas before the fields. It just gave me back the exact same SQL and said done!
There is a guy at work who has been banging on about how great it will be for doing analytics etc.
Told him 2 years ago "Don't expect much just yet. It won't solve complex stuff" but he kept blathering on - mainly because he's a shit analyst at best.
Anyway, we're a Google client at Woolies and they've been creeping in a few of their tools via Gemini to things like Sheets and Chat etc. "Summarise this chat" for when you miss things, or email trails. Yeah great.
So he proudly got us around a table to see what the offering was. I fed in a bunch of work notes from field teams and asked it a few basic questions.
While it saved me about 15 minutes writing regexp statements in BigQuery, it was hardly mind blowing.
At this stage I can see how it helps cut a few tasks down, but we're hardly talking about J.A.R.V.I.S. here...
For many people it is very useful to sort out semi-structured data, and answer the first question so you can revise and then figure out what question to ask next. Like building pylons for a bridge, I guess.
-
Scale it out to the models that take a million tokens. Now put in three years of work notes, and you can ask for trends or data points that might not be as obvious.
Absolutely has limitations, but the compute being thrown at these things is increasing 10x every six months. Elon is worried about electricity shortages…
-
I'm seeing a lot of conjecture that OpenAI have got to the point of recursive improvement, i.e. the AI is now capable of automating its own R&D and progressing towards AGI.
If correct, that's held to the timelines I read about over a decade ago which suggests those people had awfully good crystal balls. And what's concerning is those same people are genuinely worried about the impact on us.
-
@antipodean said in Chat GPT:
I'm seeing a lot of conjecture that OpenAI have got to the point of recursive improvement, i.e. the AI is now capable of automating its own R&D and progressing towards AGI.
If correct, that's held to the timelines I read about over a decade ago which suggests those people had awfully good crystal balls. And what's concerning is those same people are genuinely worried about the impact on us.
Anthropic are rumoured to have done it as well, and Zuck hinted at it too recently.
We'll see if it's true if those companies stop advertising for mid-level developers for websites.
-
AI agents are already hitting things.
Mate quit his job and started up a new business. He needed to learn some new software - eg video editing. There are loads of agents/bots trained on the software that provide realtime, useful support to actually 'do' stuff. You just ask 'how to import this type of video' and it tells you what to do. 'how to slice it' and it finds the right tool, tells you how to use it - and doesn't bombard you with all the other related tools you don't need.
Looks a game changer for learning new things - like having an expert on call to help you 24/7
-
Anyone have opinions of the strengths and weaknesses of Grok? (it is now free to use)
Had an argument with Grok yesterday. Thought its value judgement pretty horrifying, I was pushing it on the comparisons of group-identity vs individualist movements. Even when comparing death counts (100m vs potentially hundreds of people, I asked Grok to compile all the figures itself) it couldn't consider group identity groups as being negative for society. Although it was on-board with Nazism and the KKK being extremely evil, Communism and the corresponding millions of deaths wasn't as bad because of progress in 'womens rights' and the unintentional consequences of starving 10s of millions of people to death.
Tried it with Grok as I figured it was probably the least likely to have an extreme 'woke' bent to it. But AI coming out now after the last decade of nonsense is a pretty bad mix, unless it is of course just a bad ai. Would be interesting to run comparisons.
-
well well well, be curious to see how deepseek changes a few things, a bit slower as based on reinforcement learning as opposed to huge amount of data, but cost effective in comparison.
Superannuation funds probably the biggest thing hit today
Coconote
