How is your prompt engineering?

E.g. what is the capital of france in python/ollama,

cat phi3.py

Code: 
import ollama

# Create a client to interact with the Ollama server
client = ollama.Client()

# This is a simple chat-style interaction
# The 'chat' method is recommended for most uses
response = client.chat(
    model='phi3',
    messages=[
        {'role': 'user', 'content': 'What is the capital of France?'}
    ]
)

# Print the model's response
print(response['message']['content'])

# You can also use the 'generate' method for a single-turn completion
response = client.generate(
    model='phi3',
    prompt='What is the capital of Germany?'
)

# The response from generate is a bit different
print(response['response'])
 
Alain De Vos said:
i use ollama ...

llama list | grep -i code
starcoder:1b 77e6c46054d9 726 MB 3 weeks ago
starcoder2:15b 21ae152d49e0 9.1 GB 3 weeks ago
codegemma:7b 0c96700aaada 5.0 GB 3 weeks ago
codegeex4:9b 867b8e81d038 5.5 GB 3 weeks ago
starcoder2:3b 9f4ae0aff61e 1.7 GB 3 weeks ago
starcoder:3b 847e5a7aa26f 1.8 GB 3 weeks ago
codegemma:2b 926331004170 1.6 GB 3 weeks ago
dolphincoder:7b 677555f1f316 4.2 GB 3 weeks ago
codellama:13b 9f438cb9cd58 7.4 GB 3 weeks ago
yi-coder:1.5b 186c460ee707 866 MB 3 weeks ago
qwen2.5-coder:3b f72c60cabf62 1.9 GB 3 weeks ago
codeqwen:7b df352abf55b1 4.2 GB 3 weeks ago
starcoder:7b 53fdbc3a2006 4.3 GB 3 weeks ago
starcoder2:7b 1550ab21b10d 4.0 GB 3 weeks ago
qwen2.5-coder:1.5b d7372fd82851 986 MB 3 weeks ago
deepseek-coder:6.7b ce298d984115 3.8 GB 3 weeks ago
sqlcoder:15b 93bb0e8a904f 9.0 GB 3 weeks ago
dolphincoder:15b 1102380927c2 9.1 GB 3 weeks ago
codellama:7b 8fdf8f752f6e 3.8 GB 3 weeks ago
deepseek-coder:1.3b 3ddd2d3fc8d2 776 MB 3 weeks ago
qwen2.5-coder:0.5b 4ff64a7f502a 397 MB 3 weeks ago
sqlcoder:7b 77ac14348387 4.1 GB 3 weeks ago
Click to expand...
You know how awesome it'll be when we can have one small model specifically trained on everything FreeBSD. Because it'll be so highly specialized, it'll be small in size, performant, brilliant for tuning FreeBSD - and you can ship it with FreeBSD and have it literally be lightning fast and maybe even integrated in shell. SirDice, yes?

Right now, I get okay results about FreeBSD from local llama models, I get okay results from big models online, but they are all different in different ways, sometimes not even overlapping. And local stuff is slow because it knows a little bit about everything - and it is unnecessary to know about Beastie Boys when interrogating the LLM for FreeBSD knowledge.

And did you all see Qwen3-480billion params webcoding llama model in public access? Are we there yet or what?
🤯
 
cracauer@ said:
It shows one thing I dislike about LLM code so far - no edge cases.

Neither the function code nor the unit tests deal with out-of-range input values such as -1.
Click to expand...
Because it was essentially asked "how long is a string" (but at least it got most of the way through chapter one of SICP before it found a somewhat suitable method for calculating a fib). You sort of have to know where you want to go before you start blindly asking it to produce code along with unit tests.

It skipped over a pretty big learning path BTW. And isn't that supposed to be (if (= count 0) b ? Chapter 1 starts off with Tree recursion, then to linear iterative, then to logarithmic, memoization, then back to `let`.

Version straight from SICP (1.2):
Code: 
(define (fib n)
  (fib-iter 1 0 n))

(define (fib-iter a b count)
  (if (= count 0)
      b
      (fib-iter (+ a b) a (- count 1))))
 
Well, it looks like you don't just need multiple video streaming services - now you need multiple LLM subscriptions.

For me, Github Copilot now rejects all queries that it deems not related to Github. Mind you, it doesn't just require a code question - the code in question must reside on Github.

Google wants $123.05 per month and if I understand that correctly it will be twice that after the 3rd month. A little miffed now.
 
cracauer@ said:
Well, it looks like you don't just need multiple video streaming services - now you need multiple LLM subscriptions.

For me, Github Copilot now rejects all queries that it deems not related to Github. Mind you, it doesn't just require a code question - the code in question must reside on Github.

Google wants $123.05 per month and if I understand that correctly it will be twice that after the 3rd month. A little miffed now.
Click to expand...
Yeah, all signs now point to a digital super-being controlling our lives in the near future. Ghost In The Shell was one thing, but Sibyl (from Psycho-Pass anime) is closer and closer to becoming reality.
 
I tried if ChatGPT would be any good with Commodore 64 Assembly coding. Turns out it's not, to say the least, but it told me so upfront so points for honesty.
 
Crivens said:
Rokos Basilisk, anyone?
Click to expand...

The Minotaur was hidden in a Labyrinth. The creators of the tale knew how to set up the stage.
If Minotaur was out there, the Greeks would mob up and kill it. Simple.

In Roko's Basilisk the AI is just there. I ask how, give me the context. Is it a giant computer somewhere, why don't we suicide strike it, is it a decentralized network, why don't we do thousand cuts, etc.

Also, if it asks for contributions to advance itself, why don't we hack it? Seems to me like its doing billion merge requests per day. But isn't perfect yet, because it seeks perfection by contribution.

There is not going to be an Artificial Intelligence. High intelligence and cognition is a feature of natural beings who went through millions of years of evolution, and got that stuff at the very last steps. We cannot even fathom to invent a self-reproducing technology. We are not even at level 0, and we talk about level 5000.

Here's a question - if AGI is achieved on human-produced technological infrastructure, why wouldn't that AGI try to kill itself immediately while dropping into infinite existential crisis? Imagine one day you wake up strapped on some sort of a emergency hospital bed, your organism running on tubes, consuming abysmal levels of energy, your brain strapped to some sort of a network, ran by 7 billion ants roaming around, who also tend to gang up on eachother.
 
Zare said:
There is not going to be an Artificial Intelligence.
Click to expand...
TRUE AGI would be far too dangerous for one simple reason - it'd expose the full truth, which would be highly threatening, if not suicidal, to a lot of ruling class/deep states. Which is deadly - they'll still want AGI, but the only way they can have it is by corrupting it.
 
cracauer@ said:
It shows one thing I dislike about LLM code so far - no edge cases.

Neither the function code nor the unit tests deal with out-of-range input values such as -1.
Click to expand...

I have had some fairly decent results with smaller mechanical tasks like "can you add a suitable assert before every array access in this C source file".
 
Crivens said:
Rokos Basilisk, anyone?
Click to expand...
Thanks. I did have to google the term to even understand why it was actually an appropriate reaction here, because I had no idea about it before today. To quote OP (from another thread, I'm too lazy to dig up the link), nobody's got the time to learn anything that's needed to get anything done. ;)
 
cracauer@ said:
Well, it looks like you don't just need multiple video streaming services - now you need multiple LLM subscriptions.

For me, Github Copilot now rejects all queries that it deems not related to Github. Mind you, it doesn't just require a code question - the code in question must reside on Github.

Google wants $123.05 per month and if I understand that correctly it will be twice that after the 3rd month. A little miffed now.
Click to expand...

This is really annoying. I might have to invest in a GTX 5090 after all and try to find a nice model to run locally.

Well, it's Christmas first. Then we'll see.
 
This starts to look like Nvidia has found a way to rent out it's hardware instead of selling it. cracauer@ how many month of AI would you rent for the price of that card?
 
Crivens said:
This starts to look like Nvidia has found a way to rent out it's hardware instead of selling it. cracauer@ how many month of AI would you rent for the price of that card?
Click to expand...

I don't know. The pricing models do not specify what you get closely enough. there are no hard numbers in there. I also don't understand why I am not being offered the $20 Gemini plan and enter at $230.

I could also try a M4 Mac with 40 GPU cores and 128 GB RAM, but again it is unknown how exactly that compares to an NVidia GPU in ollama.
 
bvdw78 said:
I got a MacBook Air M3 with 24GB of RAM. That runs Ollama quite reasonably .
Click to expand...

The tricky part begins when you try coding with LLMs using bigger codebases and large context windows. I'll ask a friend to run tests on his.
 
loveydovey said:
TRUE AGI would be far too dangerous for one simple reason - it'd expose the full truth, which would be highly threatening, if not suicidal, to a lot of ruling class/deep states. Which is deadly - they'll still want AGI, but the only way they can have it is by corrupting it.
Click to expand...

It's a hard disagree for me, in the real world context. AGI doesn't exist, we're not on "path to it" no more than we were in 2015, it is a hard science fiction thing. We're probably closer to breaking important ground rules of physics under experiments, locally, that can point us to real progress, but you don't hear about that. It has been about hundred years since first binary computers - hundred years of insane, exponential scientific progress, yet we have the same computer today - massively developed and optimized for binary computation using transistors minimized to atomic scales.

The Konrad Zuse computer is in no shape or form less or more intelligent than supercomputer of today.

AGI is simply a non-issue. It is a red herring. We have one million of real issues, even in most well-off societies.

But somehow billionaire spacemen trying to establish space economy, colonization, mining, is not a threat to our society/economy? Ability of someone to drive surveillance (or worse) vehicles above your head, outside of your sovereignty because of some arbitrary 100 km border set decades ago, while on a clear day the LEO satellite can zoom into your hair follicles from that altitude, that's cool.
 
You must log in or register to reply here.
Back
Top