The AI is taking our jerbs thread

Mesai said:
Using AWS seems to almost always be more expensive than just buying a plan, and I think they're using Opus by default now.

Opus vs Sonnet will probably also be pretty situational, although I think Opus will more often than not, be the more expensive option. For certain tasks, it might use fewer tokens as it's a better model, but needs to be meaningful to actual get it under Sonnet, which might use more tokens for less cost.
Click to expand...
Oh yeah the bedrock per token usage is crazy money compared to using a plan. I just about manage to keep under the £90 limits on the Max plan, have occasionally needed to go up to then £180 one.

Probably gives you a reasonable idea of the loss-leading that's going with these providers. On a busy week, or when I'm doing lots of heavy refactoring, I'd be burning thousands of dollars worth of tokens a day.
 
mid_gen said:
Oh yeah the bedrock per token usage is crazy money compared to using a plan. I just about manage to keep under the £90 limits on the Max plan, have occasionally needed to go up to then £180 one.

Probably gives you a reasonable idea of the loss-leading that's going with these providers. On a busy week, or when I'm doing lots of heavy refactoring, I'd be burning thousands of dollars worth of tokens a day.
Click to expand...
Yea, would love some insight into costs for these providers. End of the year, Claude set Opus as the default model and generally you can easily use 5x+ value on a plan vs raw Bedrock. Lots of investor money I guess.
 
I've not considered the cost as I only use the one at work. But there we are looking to put some stats on the cost of seat licence over how many are using it.

Those that are paying for it, it's that for personal use, or business.
 
Realised yesterday you can take photos of your house and have AI remove all the furniture from rooms quite convincingly (depending on photo quality and amount of clutter) with minor corrections.

Aren't there people doing that the long way for estate agents? They are next for being AIed lol
 
Last edited:
Nasher said:
Realised yesterday you can take photos of your house and have AI remove all the furniture from rooms quite convincingly (depending on photo quality and amount of clutter) with minor corrections.

Aren't there people doing that the long way for estate agents? They are next for being AIed lol
Click to expand...

I wanted to personally bring to the attention of yourself an exceptional and rarely available opportunity which I feel may be of particular interest.

The property in question is a deceptively spacious and beautifully appointed executive-style residence, enviably located within one of the area’s most sought after residential addresses. It offers thoughtfully arranged accommodation finished to an exacting standard throughout, combining elegance, comfort, and versatility.

Internally, the accommodation briefly comprises a welcoming yet impressive entrance hall, a stunning principal reception room ideal for both entertaining and day-to-day living, and a comprehensively fitted kitchen offering an extensive range of units and preparation surfaces. To the first floor are a number of generously proportioned bedrooms served by a luxurious family bathroom suite.

Externally, the property continues to impress with a private and well maintained garden providing an excellent space for outdoor enjoyment, along with the added benefit of off street parking.

Given the location, presentation, and overall appeal, this is a property which I believe would suit a discerning purchaser such as yourself. Properties of this calibre are seldom available and early viewing is strongly recommended to avoid disappointment.

Should you wish to discuss this opportunity further, or arrange a viewing, please do not hesitate to contact myself at your convenience.
 
Just 'vibe' coding some python with maths using gemini in colab:
* arguing that points meet when Z=0 but completely missing when Y doesn't match.. then refusing saying it's right (the plots show it failing)
* it seems not to understand the concept of discrete set and that when comparing data are curves that the suggested interpolation isn't done

Not sure we're going to see massive advancements soon..
 

VL;DW - McKinsey used AI to review their hiring data to find it biased and analysing hires vs performance, found they were hiring the best git but lowest performers. (Nice way to say to your staff you're ****)

Quite how they analysed all the applicants that failed to be accepted, or were rejected on first sift, 20 years ago is another question.
 
Last edited:
NickK said:
Just 'vibe' coding some python with maths using gemini in colab:
* arguing that points meet when Z=0 but completely missing when Y doesn't match.. then refusing saying it's right (the plots show it failing)
* it seems not to understand the concept of discrete set and that when comparing data are curves that the suggested interpolation isn't done

Not sure we're going to see massive advancements soon..
Click to expand...
LLMs don’t have any real understanding of spatial stuff like this, it doesn’t understand the concepts, it just understands what words are related to other words.

Which can go along way to doing a lot of tasks, but it’s not actual intelligence.
 
mid_gen said:
LLMs don’t have any real understanding of spatial stuff like this, it doesn’t understand the concepts, it just understands what words are related to other words.
Click to expand...

Indeed. I see it as a simple surjective map (in set theory speak). Just to be clear I don't see the intelligence, just a very complex context sensitive map.

mid_gen said:
Which can go along way to doing a lot of tasks, but it’s not actual intelligence.
Click to expand...

I've finished that project with gemini and colab+python. It took about 6 re-writes from scratch to get the hang of some nuances in the 'thinking' that gemini takes.

I'm now onto a second - oddly a new/novel neural network (ie no weights), AI-assist had the ability to churn out a lot of code, test and results. It's claiming it's differentiating between patterns for pattern recognition but I'd like to go through the code to independently verify for this reason.
 
Last edited:
www.theguardian.com

US companies accused of ‘AI washing’ in citing artificial intelligence for job losses

While AI is having an impact on the workplace, experts suggest tariffs, overhiring during the pandemic and simply maximising profits may be bigger factors
www.theguardian.com www.theguardian.com

US companies accused of ‘AI washing’ in citing artificial intelligence for job losses
While AI is having an impact on the workplace, experts suggest tariffs, overhiring during the pandemic and simply maximising profits may be bigger factors
Click to expand...


Although interestingly the software sector crashed last week over fears that the AI promises will become realized
 
I've kinda reluctantly accepted that the startup I'm working at......the main front-end, which is fairly complicated now....it's now completely built by agents. I have made a couple of vain attempts to enforce some kind of 'good software engineering' patterns on it.....but the CEO has spent a lot of time building a very complicated army of agents so he can basically vibe code anything in there, and it works. It's not pretty under the hood, but it's not completely awful.....and you know what, it's good enough.

For the purposes of a *lot* of businesses, with a bit of experience and time and effort setting the pipelines, and using the latest models, they produce software that works sufficient to get the job done.

Back-end stuff that is more sensitive I still wouldn't trust completely to AI. But performance wise AI is actually very good, because you can just give it a performance testing harness and set it off making changes, benchmarking itself, and iterating, and leave it overnight chugging away and will get there. Quantifiable metrics like that are great feedback loops for LLMs.

I'm basically building AI agent workflows to do the job of accountants/BAs atm.....and every new model that comes out, GPT 5.3 and Opus 4.6 most recently, they get closer and closer to not needing any human intervention to get fully accurate results out.
 
Mesai said:
How are you validating that though? Trust in the results is basically the number 1 issue with LLMs and is far from being solved.
Click to expand...
Most of the output is deterministically verifiable, although one aspect needs accountants to verify, which I'm implementing the scoring framework for so we can start building a data set to fine tune with (probably Mistral Large 3).
 
mid_gen said:
Probably gives you a reasonable idea of the loss-leading that's going with these providers. On a busy week, or when I'm doing lots of heavy refactoring, I'd be burning thousands of dollars worth of tokens a day.
Click to expand...

Yeah at an individual level sure, but inference itself is a very profitable area for them (especially via the API), obviously for OAI or Anthropic there will be a subset of heavy users in the top subscription tiers who are a loss individually but they're easily netted off by the less heavy users in that tier too so those plans surprisingly still work out as being profitable overall for them. There's plenty of optimisation tricks/caching of responses behind the scenes too - essentially the cost of a given user's usage on one of those plans is way lower than the API prices would suggest. They do have an issue with using those subscription plans with third party apps, especially for agents as that just sends a load more heavy users into the given tier when those users should be paying full whack via the API.

If they didn't have more research/training runs to do and were just offering inference then they'd be printing money at an insane rate right now.
 
You must log in or register to reply here.
Back
Top Bottom