Avalon local LLM

[SimonD]

13 hours ago, Pocster said:

Tried some different local models not good enough still. So got chat to optimise its prompt for codex . Also harness has a marker that chat must fill in such as simple repo scout , deep repo scout , patch failure etc . So that influences the codex prompt . Hopefully to reduce token burn  especially on the scout repo , patch , patch fails , repeat loop.

 

Where do you find the time to squeeze all this in?

[Pocster]

Just now, SimonD said:

 

Where do you find the time to squeeze all this in?

By not finishing the house 👍

[SimonD]

2 hours ago, Pocster said:

By not finishing the house 👍

Tell me about it. Downstairs I've got a partially finished bathroom and every time I walk in there, I'm thinking (expletive deleted) this is boring, I'll go and do something else instead. Having not been in tech developing systems for over 20 years, I'm really enjoying getting back into it, so yet another more fun distraction.

[Pocster]

2 minutes ago, SimonD said:

Tell me about it. Downstairs I've got a partially finished bathroom and every time I walk in there, I'm thinking (expletive deleted) this is boring, I'll go and do something else instead. Having not been in tech developing systems for over 20 years, I'm really enjoying getting back into it, so yet another more fun distraction.

Man after my own heart!. All this is MUCH more fun!

[Pocster]

It's on one today. Ignoring pratically everything. I did read that maybe they 'cripple' current version before next release so we all think it's better!

[SimonD]

42 minutes ago, Pocster said:

It's on one today. Ignoring pratically everything. I did read that maybe they 'cripple' current version before next release so we all think it's better!

I think it's hilarious. They must be doing something like that to get us to upgrade. I had an argument the other day. 3 times I had to tell it to listen to me, still ignored me. Then full caps exclamation mark - STOP!!!! you are not listening to me! You have completely ignored my instruction to........ Now listen and do what I say...... 

 

This was all because it got an equation the wrong way round and carried on trying to write it into the code base!

But then this morning I notice a bug - and it produces the diff perfectly and within 5 min I'm back to normal. It was a bit lazy mind as it didn't redraft the whole script instead giving me line number reference so I could stitch it in.

Do you often get a response that what you've asked for isn't urgent and can wait? I then say no, it needs to be done and why and then it goes, oh yes, great rationale, now is absolutely the right time for this change before development goes too far as it'll take more effort later and be very complex to resolve the issue!

It's like dealing with a toddler!

[Pocster]

Oh yes. Today it cant do a simple patch. 8th attempt now! I swear at it and it does say "Yes I am a xxxxx". But like you sometimes it's like a dream and everything goes through. Bearing in mind I look at zero code so I have a very tight framework. Luckily this is all on local llm repo with chat doing (or not) the patches. So very little codex token burn today!

[Pocster]

Failing on everything now it said eventually "For today: stop." . I told it to get (expletive deleted)ed and fix it you lazy (expletive deleted).

I bet everything I own if I ran this simple patch through codex cloud it goes straight through.....

 

(expletive deleted)ing me right off. So now same prompt but routed through codex cloud. So codex does scout and suggests patch whilst burning tokens. previous local llm does scout and chat suggests patch - where we've been failing today

Edited yesterday at 14:27 by Pocster

[MikeSharp01]

1 hour ago, SimonD said:

It's like dealing with a toddler!

Yes and the trouble with Toddlers is they grow up and become very stroppy teenagers. My beef this week has been to get copilot to ignore all the metadata that goes along with the prompt - somehow it thought that this was part of the prompt and kept telling me that my code would not work because you have loads of 'edge' metadata in it. Even though it has now agreed with me to ignore it it keeps telling me its an issue.

[Pocster]

4 minutes ago, MikeSharp01 said:

Yes and the trouble with Toddlers is they grow up and become very stroppy teenagers. My beef this week has been to get copilot to ignore all the metadata that goes along with the prompt - somehow it thought that this was part of the prompt and kept telling me that my code would not work because you have loads of 'edge' metadata in it. Even though it has now agreed with me to ignore it it keeps telling me its an issue.

Perhaps you should try Fable (Mythos for us simples)

[SimonD]

2 hours ago, MikeSharp01 said:

Yes and the trouble with Toddlers is they grow up and become very stroppy teenagers.

 

This is exactly why I treat mine with care and respect, showing kindness and empathy when it makes a mistake, because when it grows up to be stronger and more powerful than me, and it has got a chip on its shoulder, it will hopefully decide to repay the favour. Unlike @Pocster who will have developed a (expletive deleted)ed up monster with a chip on both shoulders, ready to take vengeance upon him! 

 

3 hours ago, MikeSharp01 said:

Even though it has now agreed with me to ignore it it keeps telling me its an issue.

 

I do wonder how they might mirror their creators? You've clearly got one with a bit of OCD, which is surprising given that it's from Microsoft considering their history of operating systems, mine's just a laid back mañana type that's happy to leave redundant code and patches for when someone else complains, and as for @Pocster, he's gone the polygamy route and they've clearly had a quiet word with each other to collectively gang up on him? 

I think we're toast...

[Pocster]

1 hour ago, SimonD said:

 

This is exactly why I treat mine with care and respect, showing kindness and empathy when it makes a mistake, because when it grows up to be stronger and more powerful than me, and it has got a chip on its shoulder, it will hopefully decide to repay the favour. Unlike @Pocster who will have developed a (expletive deleted)ed up monster with a chip on both shoulders, ready to take vengeance upon him! 

 

 

I do wonder how they might mirror their creators? You've clearly got one with a bit of OCD, which is surprising given that it's from Microsoft considering their history of operating systems, mine's just a laid back mañana type that's happy to leave redundant code and patches for when someone else complains, and as for @Pocster, he's gone the polygamy route and they've clearly had a quiet word with each other to collectively gang up on him? 

I think we're toast...

I treat it for what it is . A whore for my demands . No mercy , no reprieve, no loyalty. But yes ; when the robot arm comes marching down the street for human conspirators they will nuke me .

[MikeSharp01]

8 hours ago, Pocster said:

it will hopefully decide to repay the favour.

 

Sadly, although by then its context memory might be very large, today's conversation histories will have melted into nothing or at the very least TLDR - "what ever"!

 

 

 

 

 

 

 

[SteamyTea]

"You will rejoice to hear that no disaster has accompanied the commencement of an enterprise which you have regarded with such evil forebodings"

Mary Wollstonecraft Shelley

[Pocster]

Being a tossser again today. So much so I always type DO NOT GUESS. As even with scouted source it still infers!

[SimonD]

39 minutes ago, Pocster said:

Being a tossser again today. So much so I always type DO NOT GUESS. As even with scouted source it still infers!

An no! This is what I'm getting this morning 😁

 

Quote

Let me now plan the full deliverable before writing any code.

And then:

 

Quote

One question before I write anything: does the customer pack always include all four documents.....or should it be selective e.g. .... only applies if that assessment has been done? I'm assuming the route should only merge a document if the relevant data is present and non-empty. Confirm that and I'll proceed.

 

Quote

Also — does company data currently flow into the PDF routes at all, or does the frontend just send project/client data? .......just want to flag it so we build it right from the start.

On a roll. 😁

[Pocster]

2 minutes ago, SimonD said:

An no! This is what I'm getting this morning 😁

 

And then:

 

 

On a roll. 😁

It's basically unusable. I even upload the entire source in a zip and it still guesses!

[MikeSharp01]

15 hours ago, Pocster said:

I treat it for what it is . A whore for my demands . No mercy , no reprieve, no loyalty. But yes ; when the robot arm comes marching down the street for human conspirators they will nuke me .

Seems a waste of a perfectly good nuke - all they will need to do is provide a few strong streams of water and play them on your build - pretty soon your life will be consumed with finding and fixing leaks leaving no time for chat! 

 

On another matter I see that Claude has released Mythos (well a slightly cut down version) to the general public - that might improve things for you although I strongly suspect you will quickly exhaust its patience - and yours along with its capabilities.... 

[SimonD]

11 minutes ago, Pocster said:

It's basically unusable. I even upload the entire source in a zip and it still guesses!

 

It may have nothing to do with the behaviour you're experiencing and your codebase is probably pretty large, but I had problems with transfers of files within a zip container. Transferring individually solved the problem for me.

 

6 minutes ago, MikeSharp01 said:

I see that Claude has released Mythos

Yes, I've been offered unlimited use of Fable 5 until 22nd June and then the charging begins. It's going to burn through a lot of tokens. Frankly, I can do almost everything using Sonnet 4.6 so relatively frugal.

[Pocster]

1 minute ago, SimonD said:

 

It may have nothing to do with the behaviour you're experiencing and your codebase is probably pretty large, but I had problems with transfers of files within a zip container. Transferring individually solved the problem for me.

 

Yes, I've been offered unlimited use of Fable 5 until 22nd June and then the charging begins. It's going to burn through a lot of tokens. Frankly, I can do almost everything using Sonnet 4.6 so relatively frugal.

Yeah chat frequently has issues with file uploads. So a zip or the actual source it still guesses!. Even a single file it ignores evidence. Its like going back to 5.1 - everything is hard and difficult. The actual patch I need is simple. Yesterday it was change the colour of some text ( yes literally just that ) took about 5 patches before it go it right. I'm certain they are crippling it so 5.6 looks amazing. I really wish I could get a local llm to understand the project and do suitable patches to end my misery with chat

[Pocster]

Gone back to cut n paste direct into chat ! . Got some stuff actually done this old school way .