Avalon local LLM

[Pocster]

Gone back a bit to codex as scout and patch suggester ! . Chat validates if patch is good . Codex tends to wander a bit but when it works it works ! . Saves a bit of time . Still have pi harness and local 120b as fallback .

More impressive was minstrel ! Understanding properly implied time  “ play Coldplay from yesterday “ , “ play Coldplay album I played last Tuesday “ , “ play album I had on last evening “ . Not linked to Postgres yet nor lms playback ( chat refuses to jump testing first !!! ) . Minstrel also extremely good at “mis hearing “ e.g whisper translate so “ play yeller by foldplay” gets converted correctly !

Edited Monday at 18:12 by Pocster

[MikeSharp01]

18 minutes ago, Pocster said:

Not linked to Postgres yet

I have been an SQL server & MYSQL user until about a month ago when one of my students was showing me some work they were doing on a knowledge hub with Postgres and I thought I should give it a go for my next mini project idea and I found it very good all kinds of ways the pgAdmin software is excellent.

[SimonD]

10 hours ago, MikeSharp01 said:

I have been an SQL server & MYSQL user until about a month ago when one of my students was showing me some work they were doing on a knowledge hub with Postgres and I thought I should give it a go for my next mini project idea and I found it very good all kinds of ways the pgAdmin software is excellent.

Same here, plus Sqlite and Maria DB. Open Heat Loss has been my first project using Postgres and it's great, especially the way cloud hosting providers like Railway support it natively - makes life a breeze. Although for me the only thing has been learning its quirks at the application layer so had a few issues there - but that's just learning any new DBMS.  

[Pocster]

1 hour ago, SimonD said:

Same here, plus Sqlite and Maria DB. Open Heat Loss has been my first project using Postgres and it's great, especially the way cloud hosting providers like Railway support it natively - makes life a breeze. Although for me the only thing has been learning its quirks at the application layer so had a few issues there - but that's just learning any new DBMS.  

I’m not a DB person as it’s boring stuff . Only using Postgres as chat suggested it . If it works and I haven’t got to look at it I’m happy 😃 

[SimonD]

1 hour ago, Pocster said:

I’m not a DB person as it’s boring stuff . Only using Postgres as chat suggested it . If it works and I haven’t got to look at it I’m happy 😃 

Boring yes, but also essential. The problem is that you might just end up with shite database design that's poorly normalised as I've found that AI is typically lazy when it comes to entity relationship modelling, preferring to chuck stuff in as JSON without much thought instead of distinct columns and rows, and also needs to be told when to properly use one-many & many-to-many relationships. This is because it can't deal with the full context of a development and it has to be fed smaller components and steps. I made the mistake letting it create my DB on its own to begin with and then we had to have words, after which it created a proper entity relationship diagram and model that I could review and approve. Unfortunately because I didn't feed in my DB design in at the start, we're up to something like DB migration 017 (in Postgres but about 11 more on previous DBMS as well as a full DB re-write) which wouldn't have happened if I'd known all this to begin with. At least it will write all the migration scripts, but not without a handful of crashes on deployment.

This is kind of why I'm still surprised you haven't had to touch any code at all with your project.

Edited Tuesday at 10:26 by SimonD

[Pocster]

15 minutes ago, SimonD said:

This is kind of why I'm still surprised you haven't had to touch any code at all with your project.

I’m surprised also . If DB is shit ( or indeed anything else ) I assume chat 5.6 will sort out 5.5 mess ! 😂

[MikeSharp01]

9 hours ago, Pocster said:

If DB is shit ( or indeed anything else ) I assume chat 5.6 will sort out 5.5 mess !

Ah, perhaps not a sensible assumption - it seems that your GPT (LLM) is not telling you about the self attention (is all you need - where all these machines spawn their approach from) 'quadratic bottle neck' it is trying to manage in helping you. If you double the context you effectively quadruple the compute needed to handle it so stuffing your context window with the problems you created with the last stuffing is effectively melting the planet and means that 5.6 won't sort a 5.5 generated mess because the whole thing is an exothermic reaction of sorts (compute runaway) . It creates problems and in trying to sort them it creates 4 times the problem unless you can tail the context back at every step - not easy because you a creating something which is growing. You will have already found that bigger contexts have interesting effects on the attention between tokens, it starts to weaken the connections between them - the LLM effectively has a weaker grip on the whole context. You can help by reiterating important stuff in the context / prompts (teachers will know all about the power or reiteration in learning situations). You can read more here. 

 

[Pocster]

10 hours ago, MikeSharp01 said:

Ah, perhaps not a sensible assumption - it seems that your GPT (LLM) is not telling you about the self attention (is all you need - where all these machines spawn their approach from) 'quadratic bottle neck' it is trying to manage in helping you. If you double the context you effectively quadruple the compute needed to handle it so stuffing your context window with the problems you created with the last stuffing is effectively melting the planet and means that 5.6 won't sort a 5.5 generated mess because the whole thing is an exothermic reaction of sorts (compute runaway) . It creates problems and in trying to sort them it creates 4 times the problem unless you can tail the context back at every step - not easy because you a creating something which is growing. You will have already found that bigger contexts have interesting effects on the attention between tokens, it starts to weaken the connections between them - the LLM effectively has a weaker grip on the whole context. You can help by reiterating important stuff in the context / prompts (teachers will know all about the power or reiteration in learning situations). You can read more here. 

 

Maybe 5.7 ? 🤔or 5.8 ? . Definitely 6.0 “ ChatGPT now fixes previous chat code “ - yeah - that’ll do .

[Pocster]

Chat absolutely on fire today !

SO fast ! . Responses in seconds rather than minutes . File downloads happen within 10 seconds . Blazing ! . By-passed codex and my menu today as we were only dealing with same 4 files . Significant progress ! . 2nd uma8 really helps even with background radio on as we get 2 whisper transcriptions .

[SteamyTea]

14 hours ago, Pocster said:

SO fast ! . Responses in seconds rather than minutes

Are you sure, you started this topic on the 15^th February 2026.

That is 110 days, or 2640 hours, or 158400 minutes or 9504000 seconds.

 

Coldplay have been on tour 15 times over the same period.

 

[Pocster]

26 minutes ago, SteamyTea said:

Are you sure, you started this topic on the 15^th February 2026.

That is 110 days, or 2640 hours, or 158400 minutes or 9504000 seconds.

 

Coldplay have been on tour 15 times over the same period.

 

I’m sure . Chat yesterday was the fastest I’ve ever experienced it . Whether it’s the same today or not I’ll find out later .

[Pocster]

chat back to bollocks today as it recreates half the IT infra structure in the world.

[SimonD]

49 minutes ago, Pocster said:

chat back to bollocks today as it recreates half the IT infra structure in the world.

My son was telling me about all the flaws he's found using AI and I said that wasn't it great that AI has nothing like human flaws where we forget, overlook, turn up to work hung over and useless, sometimes just totally useless without an excuse, plain obstinate, and full of bullshit and with a tendency to hallucinate all manner of things, even without drugs. You could wonder who'd designed AI in the first place 😉 It's like that ubiquitous warning in all AI UIs telling us that it's AI generated and can get things wrong, so obviously unlike humans too.  And surely the IT infrastructure definitely needs to be recreated....without doubt 😁 

What's more worrying for me is how I'm now starting to recognise which AI has written the spammy sales emails I keep getting through from companies telling me how much they can transform my business with a few clicks of a mouse.

Edited 1 hour ago by SimonD

[Pocster]

2 minutes ago, SimonD said:

My son was telling me about all the flaws he's found using AI and I said that wasn't it great that AI has nothing like human flaws where we forget, overlook, turn up to work hung over and useless, sometimes just totally useless without an excuse, plain obstinate, and full of bullshit and with a tendency to hallucinate all manner of things, even without drugs. You could wonder who'd designed AI in the first place 😉 It's like that ubiquitous warning in all AI UIs telling us that it's AI generated ad can get things wrong, so obviously unlike humans too.  And surely the IT infrastructure definitely needs to be recreated....without doubt 😁 

What's more worrying for me is how I'm now starting to recognise which AI has written the spammy sales emails I keep getting through from companies telling me how much they can transform my business with a few clicks of a mouse.

right bast today. been dragged down many a rabbit hole of things it created and patches to fix patches to fix patches to fix patches - which are (expletive deleted)ed

[SimonD]

4 minutes ago, Pocster said:

right bast today. been dragged down many a rabbit hole of things it created and patches to fix patches to fix patches to fix patches - which are (expletive deleted)ed

 

Oh, I've been there. Being stuck in that doom loop makes you want to chuck the machine out of the window. Claude, however, as been good to me today. Sensibly it's reminded me of design decisions I made a while back to pull me back from a minor tantrum I was having with a bug and database versus JSON. It's just pulled out a quote from the backlog to bring me back. Kept me on the straight and narrow. Thank goodness for a good backlog where I've asked it to log design decisions!

[Pocster]

48 minutes ago, SimonD said:

 

Oh, I've been there. Being stuck in that doom loop makes you want to chuck the machine out of the window. Claude, however, as been good to me today. Sensibly it's reminded me of design decisions I made a while back to pull me back from a minor tantrum I was having with a bug and database versus JSON. It's just pulled out a quote from the backlog to bring me back. Kept me on the straight and narrow. Thank goodness for a good backlog where I've asked it to log design decisions!

I tell chat it’s a useless (expletive deleted) and it replies “ I am a useless (expletive deleted) and you are right to be pissed at me “ . I then tell it “ grow a vag and fix it “ . It’s a love hate relationship .

What’s bizarre is it’s not about the code anymore it’s about doing it correctly. After all I’m angry at its method / logic not its code because in under 30 seconds that can be fixed . It’s all a bit strange still 

Edited just now by Pocster