UPDATE: Reviews section refreshed, redesigned, searchable: Go take a look
Take the example earlier of the rounding error. That makes perfect sense for a pure LLM. It's not doing any maths, just trying to predict the most likely next token. So, if it's seen the string "7.393" near to the string "7.3874" and the string "rounded" in its training data then it may say that's the most likely result. But I think you can now tell most frontier models to look for maths and when it finds an equation to generate a bit of code to solve it rather than just using the LLM.
From the maths it's resolved for me so far it was a suprise to see that. Rounding is basic logic and numbers, and numbers or letters, it's just pattern matching logic as far as I understand it. It is good at finding errors i.e. when you know something isn't right (I'm using a 3D CAD model to verify the chatbot/AI/LLM output) and challenge it, the model and conversation improves.
Certainly it shows how you can't reliably use it to simply do work that you don't have some understanding of or an ability to spot errors in, but if the original Apple was a bicycle for the brain I'm finding current LLMs more like a motorbike (or an Avinox e-MTB.. ha).
The half dozen AIs i've used are saving me a colossal amount of time. They can produce a days work in minutes with a handful of decent prompts.
Unfortunately it takes me almost as long to error check and edit the output, and make sure it's not lying.
It's also bloody useless at context dependent language.
So, yeah, in an eight hour day, it saves me about 20 minutes. Colossal.
I'm so pleased about it.
Hell, it even makes mistakes summarizing a 1000 word meeting minutes, bad enough that the entire context is broken.
Sam Altman has admitted being wrong in his predictions about AI job losses.
"I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened... that is an area where my intuitions were just off.”
It hasn't made headlines but this seems pretty damning, one of the key AI chiefs admitting that his product can't really do what he's been promising afterall
My boss gets loads of AI emails, he uses AI to collate and summarise them and uses it again to write replies... He thinks AI basically talking to itself all day long is a huge benefit to the company. As a fairly plain speaking person, I found it handy to chuck all my thoughts into it and get a management level proposal out of it.
@jameso, interested to hear how you're applying AI and CAD models as while I get loads of people trying to sell me AI CAD add-ons, I'm yet to find any real useful crossover in that area.
Meanwhile,
“I’m delighted to be wrong about this, I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened,” Altman said.
I think by now is a key part of that. " he's not saying it can't do what he said, just that it hasn't done it yet.
“I’m delighted to be wrong about this, I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened,” Altman said.
I think by now is a key part of that. " he's not saying it can't do what he said, just that it hasn't done it yet.
That's true, he does leave open the possibility that it might deliver in the future. But I'm not sure he has much choice, considering he still needs investors pumping money in, and they aren't going to do that if he openly admits that the tech isn't capable of living up to the promises.
I think there's a good chance that this is one of those things that's always "2-3 years away" from delivering. The AI CEO's have been very aggressive with their predictions but their initial timelines are running out and there's no real sign of the jobs apocalypse, and this is the first time I've seen them strike a more cautious tone so it's an interesting development.
It'll be fascinating to see how this plays out over the next 12 months. What happens when an industry which has attracted hundreds of billions in investment simply fails to deliver on its promises, I don't know.
I do not get the same result What am I missing about that image?
The explanation at the bottom of it?
99% of the time I avoid deliberate AI use, although I notice you get it with google search and that does lead to deliberate prompt writing rather than thinking of search terms. Sometimes my curiosity gets the better of me and I deliberately use it, to try and help me understand a topic I find difficult to understand (ie math/code problems - hobby level).
I don't think it's brilliant at all. The costs to the environment, through the size of the data centres, to their massive water and energy requirements - especially at a time of climate upheavel with wildfires - are enough for me to not want to use it and certainly won't be cheering it on.
Got ai stuff switched off on my iPhone 17/ipad pro etc - no need for it myself but LLM are not what I’d consider to be ai anyway
Got ai stuff switched off on my iPhone 17/ipad pro etc - no need for it myself but LLM are not what I’d consider to be ai anyway
I switched it on (iPhone 15pro), after it had taken up 5GB of space and was up and running I gave it a week’s use. I found it added nothing to the experience of using the phone and turned it off, regaining those 5GB of storage for other stuff.
Still use it a lot, but just for what it is - a very big word processor. It's too backward looking, and doesn't catch or is able to work out human traits, such as lying. Which renders it pointless for enforcement.
@jameso, interested to hear how you're applying AI and CAD models as while I get loads of people trying to sell me AI CAD add-ons, I'm yet to find any real useful crossover in that area.
Sorry, missed this earlier.
I'm not running it actually within CAD so prob not what you're looking to do or being sold. My CAD level is quite simple, detailed within a narrow area that's all. What I've found AI excellent for is building simple apps that use maths differently to CAD's timeline process. In CAD I can only drive one element via something that went before it, but some systems are interdependent in a loop and need plotting to find locus points. I found AI handy to produce the code for a mini-app that either gives the numbers needed for the model, or more simply verifies the CAD model at a given point, just check a few points to know it's tracking as it should. (edit to add, there's got to be ways to link that to the parameters table better but it's still in that timeline structure).
I'm finding a lot of value in AI, tbh. I had a question in mind about bike geometry maybe 10 years ago now and was only able to answer it for sure 6 months ago via AI/LLM use. In the process of trying to figure it out I found some related maths and realised how incredibly smart some people are, so for a while I gave up on it bc I simply could never use maths at that level. Then I got further with the problem using CAD and it looked about right but I couldn't verify it and I knew the CAD model was simplified because of that chronological point. The AI tool gave me a checker that backed up the CAD model and bingo, I have the answer. Nerdy stuff but very satisfying in a 'not many people know this' kind of way.
I think AI may be over-inflated in value now and we got distracted by critiquing the creative output from it, but that's not where it has as much value and anyone using it substantially for work like that is in a cul de sac. After answering that question I realised how LLMs are 'a bicycle for the brain' in the same way the first PCs were. I've used it to make a couple of really useful tools for work and personal interest. Would I pay for a brain-boosting tool like this? Of course, and there is its value, as well as the risk if an access gap opens up.
When all you have is a hammer, everything looks like a nail?
Of late I've been using ChatGPT for programming. It is quite remarkably good. It puked out a Python program in seconds which would have taken me - still learning the language - days. It then sat there critiquing its own work and suggesting performance enhancements using features I've never heard of.
Even if you take the default stance of "it probably writes poor code and you can't trust it, you'll have to check its work," it's an amazing teaching aid. I learned more Python in a week than I have in months of working from reference guides and tutorials.
I had a similar experience to Cougar. The last time I took a programming class was 1986 learning to code in Basic and saving our few dozen lines of code to cassette tape. That persuaded me I had no interest in working as a programmer.
An NPO that I help with some technical analysis shifted to a new data collection system, but the database is not the relational database that the software we use for the analysis requires. The server admin wrote some Python code that exported the required data, but he didn't understand the analysis so it was very basic and still needed a lot of manual work to prepare it for analysis. I gave his code to Copilot and it produced much more useful code after about 15 minutes of prompting. It still wasn't quite right though, so I gave it to Gemini and asked it to fix it. That was an improvement, so I gave it back to Copilot, and so on. After a few rounds of this, it did everything I wanted. Even a novice could use the script to export the data, run the analysis, and export the basic results.
After that, I had it code a couple of more Python scripts, one of which is basically a matchmaking system (for example, like you would expect in an online dating system) that outputs the results in a format that mail merge systems can read.
The second was much more ambitious, I asked it to write Python code to run the actual analysis (because the commercial software I use is available for Windows only and all the Mac fanbois on the project refuse to use Windows, so I'm the only person who can run the analysis.) The analysis is multidimensional (in this case 3 dimensions, but the commercial software can handle up to 99) and the three dimensions have to be calibrated to a common benchmark. The Python script could recover the three dimensions and they correlated 1.00 with the results from the commercial software, but it couldn't calibrate them to the common benchmark. In this case, it doesn't matter. It's analogous to comparing the height of trucks with the clearance of bridges. If you only care about comparing the trucks with each other, the bridges and trucks don't need to be calibrated. However, if you want to compare the trucks with the bridges, the calibration is essential. In this case, we are only interested in comparing the trucks with each other, so it's functional as a proof-of-concept for what I wanted to do, but not for general use. I spent a few days playing with it, trying to get it to calibrate, both Copilot and Gemini said they could see the problem and claimed to fix it, but never managed to. Still pretty impressive. I'm sure that someone who knows how to code save a huge amount of time with it.
Similar story. I wanted some Mac "Quick Action" scripts to do some simple image manipulation - resizing, adding borders, adding captions from the image metadata and suchlike. I could have done it myself, eventually, but it would have taken ages to learn and relearn the relevant bits of ImageMagick, exiftool and basic shell programming. ChatGPT did it in minutes.
Yeah.
The program I "wrote," I won't go into detail because it's not important but it takes in a big number and spits out a different big number. Really big numbers take a hellishly long time to compute. ChatGPT went "what you want there mate is a cache." Now the first compute still takes ages but after that first one it barks out results literally instantaneously. I wouldn't have thought of that solution in years if ever, I didn't even know it existed let alone how to code it.
TLDR - I learned something.
I use AI daily in work... Small stuff, niff naff and trivia, summarising things, concocting responses, whipping up slides or something, getting it to do first pass tech analysis, propose solutions to problems, all sorts. Everything needs checking... just like the average human employee output TBH.
Like it or not it is here to stay.
Yes, that was the interesting thing about my attempts. It explained what each part of the code was doing. Reading the code, it was mostly fairly transparent what was happening even though I'd never seen Python code before. The more specialized parts of the analysis required installing extra modules, but it explained what those did. Seems like it would be a great learning tool if used sensibly. Comparing code between different AIs was the key thing though, multiple iterations of that were needed to get the complex stuff to work properly.
Sorry, missed this earlier.
No worries. Not directly how I use AI or CAD but I can see what you're getting at. Thanks
Have any of the programmer types here used a locally hosted AI coding assistant via Ollama or similar?
Any recommendations on the model to use? Must have zero reliance on connection to the net.
i want it to do basic stuff in Python. Add type hints, write unit tests, class boilerplate etc. GitHub copilot seems to have got worse at doing this (certainly slower) at the expense of doing more complex stuff that I don’t want it to do!
I have not used any local models recently but I did see that Meta have released what looks to be just what you need.
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Yeah, I read about that and looks promising.
i just have to get over my feelings about Zuckerberg and the whole Meta / Facebook corporation. (That they’re psychopaths)
They are psychopaths! Also Zucks looks like a serial killer....
I use AI daily in work... Small stuff, niff naff and trivia, summarising things, concocting responses, whipping up slides or something, getting it to do first pass tech analysis, propose solutions to problems, all sorts. Everything needs checking... just like the average human employee output TBH.
I use it in pretty much the same way - turn this spreadsheet into something a normal human understands. Create a Power Point, write a shitty letter in a nice way. It's saving me time fo'shure.
The expense it's taken to create that still isn't sustainable though.
If you're curious about Facebook, then get hold of a copy of Careless People - it's mind boggling.
I also read We Are Not Machines by Sarah O'Connor over the summer. It's a really interesting bit of reportage (the author is an FT correspondent, so it's very well written, too) that isn't as grim about the future of work and AI as I expected it to be.
I find it depressing that so many people are just not giving enough of a shit about climate change and are able to put it to the back of their mind while stood staring directly in the face of wildfires for the sake of convenience and productivity. Burn in hell corporate swines!
Booking.Yeah
😉
But your right. We all want, bugger the cost and world.
Are we the last generation.
While I get your concern, I'd love to see how much difference AI actually makes to climate change considering we've two wars going on that seem to daily blow up more oil than I'd use in multiple lifetimes.
"Have any of the programmer types here used a locally hosted AI coding assistant via Ollama or similar?"
Yes, it wasn't very good. It was qwen or something and I linked it to Continue in vscode. It can do basic stuff fine, but limited reasoning over large files, forget large codebases of multiple files! But it still might be helpful. It was a while ago, someone must have written into a codex style tool which can search through files more efficiently with cmd line tools and use the llm on the result?
I ran it on a gaming laptop, think rtx4060 with 12gb vram
This was interesting...
https://www.bbc.co.uk/news/articles/cvgx4yd1gl2o
You'd think OpenAI and Anthropic etc wouldn't really need any AI Engineers and Software Developers?
I use it in pretty much the same way - turn this spreadsheet into something a normal human understands. Create a Power Point, write a shitty letter in a nice way. It's saving me time fo'shure.
The expense it's taken to create that still isn't sustainable though.
Agreed.
A couple of things will probably happen to the cost, pricing will be priced on the real cost of running this stuff, we're only just exiting the land grab phase, and the models will be scaled down and/or made more efficient, AFAIUI things are still basically brute forcing their way through by throwing compute power at oceans of tokens.
Have any of the programmer types here used a locally hosted AI coding assistant via Ollama or similar?
Any recommendations on the model to use? Must have zero reliance on connection to the net.
i want it to do basic stuff in Python. Add type hints, write unit tests, class boilerplate etc. GitHub copilot seems to have got worse at doing this (certainly slower) at the expense of doing more complex stuff that I don’t want it to do!
Local agents aren't really there yet, for anything more than a handful of lines or basic functionality.
Github copilot is aimed at what you're talking about, it seems as good as anything else as far as I can tell. If you're relying on the free version of chat gpt, you can definitely expect an extreme quantity of enshittification to be coming that way.
Caveats: I use it at work; the models are bought and paid for, I didn't do the analysis on the local agents.
While I get your concern, I'd love to see how much difference AI actually makes to climate change considering we've two wars going on that seem to daily blow up more oil than I'd use in multiple lifetimes.
Good point, let's not forget the military usage of AI for blowing up children in schools, but yeah productivity.
Did see recently the claim the US military is the single largest polluter on the planet, due to all the transportation they do, continually/globally.
AI prompt writing is the new SEO sector.
Also, all aboard "the oops sorry we accidentally let our AI hack some company, isn't it marvellous" boat.
Also, all aboard "the oops sorry we accidentally let our AI hack some company, isn't it marvellous" boat.
I really liked the press release from Meta a couple of days after the two big AI brands announced theirs had gone rogue saying "yeah ours did that too, honest".
I find it depressing that so many people are just not giving enough of a shit about climate change and are able to put it to the back of their mind while stood staring directly in the face of wildfires for the sake of convenience and productivity. Burn in hell corporate swines!
I agree in principle. But the CO2 cost of average personal prompts is really not that much. Less than 0.5g of CO2 for a text request and reply - if an AI-generated search reply is correct, ha. More involved use, eg. to get a HTML app that uses code from a detailed prompt may be one session, then iterate it 3-5x might be a typical follow up session - estimates seem to suggest 20g per prompt, with text to-fro being far lower. So maybe 50-60g up to 100g per session. 100g of CO2 is driving an average petrol car for about 500 metres. I might have been up to a couple of kilos for that bike geometry calculation (good chance it was a lot less but let's over-estimate for now), so it equals driving 10km. I generally don't drive those kind of distances, I cycle them.
Also, all aboard "the oops sorry we accidentally let our AI hack some company, isn't it marvellous" boat.
I really liked the press release from Meta a couple of days after the two big AI brands announced theirs had gone rogue saying "yeah ours did that too, honest".
Open AI have been making them claims for years, this is them back in 2019 saying their GPT2 model was "too dangerous to release".
I think it's fair to say that GPT2 was not, in fact, too dangerous to release.
Good to see more scepticism around these kind of claims. Personally I strongly suspect they're mainly publicity stunts aimed at keeping the investors interested, though obviously any security concerns do need to be taken seriously.
Personally I strongly suspect they're mainly publicity stunts aimed at keeping the investors interested, though obviously any security concerns do need to be taken seriously.
All their promises are bollox, it’s a giant circle jerk passing around “memorandum of intentional funding” which in normal speech means absolutely **** all but sounds good when illiterate reporters publish such crap as “blah-blah company is investing £500B+ in ai data centres”
Why do you think all these tech freaks are building bunkers/security compounds that are fully off grid in remote locations around the world - they know the crash is coming and so much money has been poured into their sham that when it does inevitably come crashing down it will make the 2008 fallout look like a tiny accounting error
https://www.bbc.co.uk/news/articles/c78gr0jv0mdo
Big Tech claims AI data centers don’t use much water.
But data centers typically report the water used inside the facilities — we found them indirectly consuming 10x that amount.
We went to the drought-stricken West, where farmers now have to fight AI oligarchs for water
This video examines why billions of dollars in enterprise AI projects are being abandoned, from hallucinating customer-service bots and fabricated legal citations to insurance companies excluding AI-related claims from coverage. We also explore why Fortune 500 companies increasingly classify AI as a material business risk even while Big Tech pours hundreds of billions into data centers and NVIDIA hardware.
The deeper problem is economics. GPU prices are collapsing, infrastructure is depreciating faster than expected, AI projects often fail to produce measurable returns, and companies are borrowing enormous amounts to keep the buildout going. At the same time, the AI systems actually succeeding tend to be narrow, measurable tools where mistakes are caught immediately.
The AI revolution isn’t disappearing, but the dream of autonomous systems replacing entire departments may already be getting quietly dismantled behind closed doors.
