Over the holidays, we published a post[1] on using high-precision few-shot examples to get `gpt-4o-mini` to perform similar to `gpt-4o`. I just re-ran that same experiment, but swapped out `gpt-4o-mini` with `phi-4`.
`phi-4` really blew me away in terms of learning from few-shots. It measured as being 97% consistent with `gpt-4o` when using high-precision few-shots! Without the few-shots, it was only 37%. That's a huge improvement!
By contrast, with few-shots it performs as well as `gpt-4o-mini` (though `gpt-4o-mini`'s baseline without few-shots was 59% – quite a bit higher than `phi-4`'s).
I like the direction, but have a pretty different experience in practice. This spans legal analytics, social media analytics, code synthesis, news analysis, cyber security LLMs, etc:
1. The only ultimate absolute quality metric I saw in that blogpost afaict was expert agreement... at 90%. All of our customers would fire us at that level across all of the diff b2b domains we work in. I'm surprised 90% is considered acceptable quality in a paying business context like retail.
2. Gpt-4o-mini is great. I find we can get, for these kind of simple tasks you describe, gpt-4o-mini to achieve about 95-98% agreement with gpt-4o by iteratively manually improving prompts over increasingly large synthetic evals. Given data and a good dev, we do this basically same-day for a lot of simple tasks, which is astounding.
I do expect automatic prompt optimizers to win here long-term, and keep hopefully revisiting dspy et al. For now, they fail over standard prompt engineering. Likewise, I do believe in example learning over time for areas like personalization.... but doing semantic search recall of high-rated answers was a V1 thing we had to rethink due to too many issues.
re: 90% – this particular case is a fairly subjective and creative task, where humans (and the LLM) are asked to follow a 22 page SOP. They've had a team of humans doing the task for 9 years, with exceptionally high variance in performance. The blended performance of the human team is meaningfully below this 90% threshold (~76%) – which speaks to the difficulty of the task.
It's, admittedly, a tough task to measure objectively though, in that it's like a code review. If a Principal Engineer pointed out 20 deficiencies in a code change and another Principal Engineer pointed out 18 of the same 20 things, but also pointed out 3 other things that the first reviewer didn't, it doesn't necessarily mean either review is wrong – they just meaningfully deviate from each other.
In this case, we chose an expert that we treat as an objective "source of truth".
re: simple tasks – We run hundreds of thousands of tasks every month with more-or-less deterministic behavior (in that, we'll reliably do it correctly a million out of a million times). We chose a particularly challenging task for the case-study though.
re: in a paying business context – FWIW, most industries are filled with humans doing tasks where the rate of perfection is far below 90%.
What's your loop for prompt engineering with GPT-4o? Do you feed the meta-prompter the misclassified examples? Also does the evaluation drive the synthetic data production almost like boosting?
This is really nice. I loved the detailed process and I'm definitely gonna use it. One nit though: I didn't understand what the graphs mean, maybe you should add the axes names.
In our case, yes we treat them the same. Though it might be interesting to decouple them.
You could, for example, include all few-shots that meet the similarity threshold, but you’ll use more tokens for (I assume) marginal gain. Definitely worth a try though.
We have, and it works great! We currently do this in production, though we use it to help us optimize for consistency between task executions (vs the linked post, which is about improving the capabilities of a model).
Phrased differently, when a task has many valid and correct conclusions, this technique allows the LLM to see "How did I do similar tasks before?" and it'll tend to solve new tasks by making similar decisions it made for previous similar tasks.
Two things to note:
- You'll typically still want to have some small epsilon where you choose to run the task without few-shots. This will help prevent mistakes from propagating forward indefinitely.
- You can have humans correct historical examples, and use their feedback to improve the large model dynamically in real-time. This is basically FSKD where the human is the "large model" and the large foundation model is the "small model".
Is anyone blown away by how fast we got to running something this powerful locally? I know it's easy to get burnt out on llms but this is pretty incredible.
I genuinely think we're only 2 years away from full custom local voice to voice llm assistants that grow with you like JOI in BR2049 and it's going to change how we think about being human and being social, and how we grow up.
I've been experimenting with running local LLMs for nearly two years now, ever since the first LLaMA release back in March 2023.
About six months ago I had mostly lost interest in them. They were fun to play around with but the quality difference between the ones I could run on my MacBook and the ones I could access via an online API felt insurmountable.
This has completely changed in the second half of 2024. The models I can run locally had a leap in quality - they feel genuinely GPT-4 class now.
They're not as good as the best hosted models (GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet) but they're definitely good enough to be extremely useful.
This started with the Qwen 2 and 2.5 series, but I also rate Llama 3.3 70B and now Phi-4 as GPT-4 class models that run on my laptop.
I'm in complete agreement with your more recent timeline piece (the negative one), and as a younger user (22 year old student) I'm actively relocating this year to somewhere slightly more rural with a focus on physical/knowledge combined work to secure a good quality of life nearly solely because of how fast our timelines are.
A 'word calculator' this effective is the best substitute that we have for a logic calculator. And the fact that it's enough in 90% of situations is terrifying as it is transformative, as is the fact no one is awake to it.
Exponential power scaling in an unstable world feels like it only makes it exponentially more unstable though.
I am blown away: a year ago I bought a M2 32G Mac to run local models. It seems like what I can run locally now just one year later is 10x more useful for NLP, data wrangling, RAG, experimenting with agents, etc.
Not related to local LLMs, but JOI from BR2049 is essentially what Replika is striving for: https://replika.com/
Infact during the onboarding process they ask the user to choose which AI companion movie they related to the most: Her, BR2049 or Ex-Machina. The experience is then tailored to align closer to the movie chosen.
It's quite a terrible app from a product design perspective: filled with dark patterns (like sending the user blurred images to "unlock*) and upsells, but it's become successful amongst the masses that have adopted it, which I find fascinating. 30m+ users https://en.wikipedia.org/wiki/Replika#:~:text=Replika%20beca....
I’ve thought for a while that Joi in BR2049 was less dystopian than what we will probably do with AI. She doesn’t constantly prompt K to buy more credits (like a mobile game) to continue engaging with her or deepen their relationship. (“If you really love me…”) I’ve been expecting that this is how our industry would operate given the customer hostile psychologically abusive hellscape of social and mobile. Of course there’s still time.
She appears to be a local model runnable on a small device without cloud.
I expect AI to be like any other tech: some fantastic uses that advance humanity and improve the world, some terrible uses that abuse, manipulate, oppress.
I don’t see anything in the tech that indicates a singular pattern that will be “good” or “bad”.
It’s odd that MS is releasing models they are competitors to OA. This reinforce the idea that there is no real strategic advantage in owning a model. I think the strategy is now offer cheap and performant infra to run the models.
> This reinforce the idea that there is no real strategic advantage in owning a model
For these models probably no. But for proprietary things that are mission critical and purpose-built (think Adobe Creative Suite) the calculus is very different.
MS, Google, Amazon all win from infra for open source models. I have no idea what game Meta is playing
I think they're commoditizing their complement [1]. Engaging content helps Meta, and LLMs make it easier to create that content. Their business model has never been selling API access and releasing the model enables the community to improve it for them.
That's exactly what they're saying: it's interesting that Microsoft came to the same conclusion that Meta did, that models are generally not worth keeping locked down. It suggests that OpenAI has a very fragile business model, given that they're wholly dependent on large providers for the infra, which is apparently the valuable part of the equation.
According to many press stories in the past year, the relationship between Microsoft and OpenAI has been very strained. It looks more and more like that both sides are looking for opportunity to jump ship.
This is a very clever move by Microsoft. OpenAI has no technological moat and a very unreliable partner.
Was disappointed in all the Phi models before this, whose benchmark results scored way better than it worked in practice, but I've been really impressed with how good Phi-4 is at just 14B. We've run it against the top 1000 most popular StackOverflow questions and it came up 3rd beating out GPT-4 and Sonnet 3.5 in our benchmarks, only behind DeepSeek v3 and WizardLM 8x22B [1]. We're using Mixtral 8x7B to grade the quality of the answers which could explain how WizardLM (based on Mixtral 8x22B) took 2nd Place.
Unfortunately I'm only getting 6 tok/s on NVidia A4000 so it's still not great for real-time queries, but luckily now that it's MIT licensed it's available on OpenRouter [2] for a great price of $0.07/$0.14M at a fast 78 tok/s.
Because it yields better results and we're able to self-host Phi-4 for free, we've replaced Mistral NeMo with it in our default models for answering new questions [3].
Interesting eval but my first reaction is "using Mixtral as a judge doesn't sound like a good idea". Have you tested how different its results are from GPT-4 as a judge (on a small scale) or how stuff like style and order can affect its judgements?
Yeah we evaluated several models for grading ~1 year ago and concluded Mixtral was the best choice for us, as it was the best model yielding the best results that we could self-host and distribute the load of grading 1.2M+ answers over several GPU Servers.
We would have liked to pick a neutral model like Gemini which was fast, reliable and low cost, unfortunately it gave too many poor answers good grades [1]. If we had to pick a new grading model now, hopefully the much improved Gemini Flash 2.0 might yield better results.
I tested Phi-4 with a Japanese functional test suite and it scored much better than prior Phis (and comparable to much larger models, basically in the top tier atm). [1]
The one red-flag w/ Phi-4 is that it's IFEval score is relatively low. IFEval has specific types of constraints (forbidden words, capitalization, etc) it tests for [2] but its one area especially worth keeping an eye out for those testing Phi-4 for themselves...
IMO SO questions is not a good evaluation. These models were likely trained on the top 1000 most popular StackOverflow questions. You'd expect them to have similar results and perform well when compared to the original answers.
For structured output from anywhere I'm finding https://github.com/BoundaryML/baml good. It's more accurate than gpt-04-mini will do on its own, and any of the other JSON schema approaches I've tried.
Yeah it's not as strong as constrained beam search like OpenAI uses (at least afaik) but it works on any models that support tool calling. Just keep it simple, don't have a lot of deep nested structures or complicated rules.
Lots of other models will work nearly as well though if you just give them a clear schema to follow and ask them to output json only, then parse it yourself. Like I've been using gemma2:9b to analyze text and output a json structure and it's nearly 100% reliable despite it being a tiny model and not supporting tools or structured output officially.
I’ve seen on the localllama subreddit that some GGUFs have bugs in them. The one recommended was by unsloth. However, I don’t know how the Ollama GGUF holds up.
Ollama can pull directly from HF, you just provide the URL and add to the end :Q8_0 (or whatever) to specify your desired quant. Bonus: use the short form url of `hf` instead of `huggingface` to shorten the model name a little in the ollama list table.
Edit: so for example of you want the unsloth "debugged" version of Phi4, you would run:
`$ollama pull hf.co/unsloth/phi-4-GGUF:Q8_0`
(check on the right side of the hf.co/unsloth/phi-4-GGUF page for the available quants)
You still need to make sure the modelfile works so this method will not run out of the box on a vision GGUF or anything with special schemas. Thats why mostly a good idea to pull from ollama directly.
Phi-4's architecture changed slightly from Phi-3.5 (it no longer uses a sliding window of 2,048 tokens [1]), causing a change in the hyperparameters (and ultimately an error at inference time for some published GGUF files on Hugging Face, since the same architecture name/identifier was re-used between the two models).
For the Phi-4 uploaded to Ollama, the hyperparameters were set to avoid the error. The error should stop occurring in the next version of Ollama [2] for imported GGUF files as well
In retrospect, a new architecture name should probably have been used entirely, instead of re-using "phi3".
Readit News
`phi-4` really blew me away in terms of learning from few-shots. It measured as being 97% consistent with `gpt-4o` when using high-precision few-shots! Without the few-shots, it was only 37%. That's a huge improvement!
By contrast, with few-shots it performs as well as `gpt-4o-mini` (though `gpt-4o-mini`'s baseline without few-shots was 59% – quite a bit higher than `phi-4`'s).
[1] https://bits.logic.inc/p/getting-gpt-4o-mini-to-perform-like
1. The only ultimate absolute quality metric I saw in that blogpost afaict was expert agreement... at 90%. All of our customers would fire us at that level across all of the diff b2b domains we work in. I'm surprised 90% is considered acceptable quality in a paying business context like retail.
2. Gpt-4o-mini is great. I find we can get, for these kind of simple tasks you describe, gpt-4o-mini to achieve about 95-98% agreement with gpt-4o by iteratively manually improving prompts over increasingly large synthetic evals. Given data and a good dev, we do this basically same-day for a lot of simple tasks, which is astounding.
I do expect automatic prompt optimizers to win here long-term, and keep hopefully revisiting dspy et al. For now, they fail over standard prompt engineering. Likewise, I do believe in example learning over time for areas like personalization.... but doing semantic search recall of high-rated answers was a V1 thing we had to rethink due to too many issues.
It's, admittedly, a tough task to measure objectively though, in that it's like a code review. If a Principal Engineer pointed out 20 deficiencies in a code change and another Principal Engineer pointed out 18 of the same 20 things, but also pointed out 3 other things that the first reviewer didn't, it doesn't necessarily mean either review is wrong – they just meaningfully deviate from each other.
In this case, we chose an expert that we treat as an objective "source of truth".
re: simple tasks – We run hundreds of thousands of tasks every month with more-or-less deterministic behavior (in that, we'll reliably do it correctly a million out of a million times). We chose a particularly challenging task for the case-study though.
re: in a paying business context – FWIW, most industries are filled with humans doing tasks where the rate of perfection is far below 90%.
search(T,θ,m) retrieves the first m historical tasks that are semantically similar above the θ threshold
Are both m's here the same or different numbers? I found this a bit confusing
You could, for example, include all few-shots that meet the similarity threshold, but you’ll use more tokens for (I assume) marginal gain. Definitely worth a try though.
Phrased differently, when a task has many valid and correct conclusions, this technique allows the LLM to see "How did I do similar tasks before?" and it'll tend to solve new tasks by making similar decisions it made for previous similar tasks.
Two things to note:
I genuinely think we're only 2 years away from full custom local voice to voice llm assistants that grow with you like JOI in BR2049 and it's going to change how we think about being human and being social, and how we grow up.
I've been experimenting with running local LLMs for nearly two years now, ever since the first LLaMA release back in March 2023.
About six months ago I had mostly lost interest in them. They were fun to play around with but the quality difference between the ones I could run on my MacBook and the ones I could access via an online API felt insurmountable.
This has completely changed in the second half of 2024. The models I can run locally had a leap in quality - they feel genuinely GPT-4 class now.
They're not as good as the best hosted models (GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet) but they're definitely good enough to be extremely useful.
This started with the Qwen 2 and 2.5 series, but I also rate Llama 3.3 70B and now Phi-4 as GPT-4 class models that run on my laptop.
I wrote more about this here: https://simonwillison.net/2024/Dec/31/llms-in-2024/#some-of-...
A 'word calculator' this effective is the best substitute that we have for a logic calculator. And the fact that it's enough in 90% of situations is terrifying as it is transformative, as is the fact no one is awake to it.
Exponential power scaling in an unstable world feels like it only makes it exponentially more unstable though.
BTW, a few days ago I published a book on using Ollama. Here is a link to read it online https://leanpub.com/ollama/read
Infact during the onboarding process they ask the user to choose which AI companion movie they related to the most: Her, BR2049 or Ex-Machina. The experience is then tailored to align closer to the movie chosen.
It's quite a terrible app from a product design perspective: filled with dark patterns (like sending the user blurred images to "unlock*) and upsells, but it's become successful amongst the masses that have adopted it, which I find fascinating. 30m+ users https://en.wikipedia.org/wiki/Replika#:~:text=Replika%20beca....
She appears to be a local model runnable on a small device without cloud.
I don’t see anything in the tech that indicates a singular pattern that will be “good” or “bad”.
Hunyuan (open source video) has been remarkable. Flux dev makes some incredible images.
The fact that it's still only going to get better from here is hard to imagine.
For these models probably no. But for proprietary things that are mission critical and purpose-built (think Adobe Creative Suite) the calculus is very different.
MS, Google, Amazon all win from infra for open source models. I have no idea what game Meta is playing
Based on their business moves in recent history, I’d guess most of them are playing Farmville.
I think they're commoditizing their complement [1]. Engaging content helps Meta, and LLMs make it easier to create that content. Their business model has never been selling API access and releasing the model enables the community to improve it for them.
[1] https://gwern.net/complement
> I think the strategy is now offer cheap and performant infra to run the models.
Is this not what microsoft is doing? What can microsoft possibly lose by releasing a model?
This is a very clever move by Microsoft. OpenAI has no technological moat and a very unreliable partner.
Yes, because you can't build a moat. Open source will very quickly catch up.
Unfortunately I'm only getting 6 tok/s on NVidia A4000 so it's still not great for real-time queries, but luckily now that it's MIT licensed it's available on OpenRouter [2] for a great price of $0.07/$0.14M at a fast 78 tok/s.
Because it yields better results and we're able to self-host Phi-4 for free, we've replaced Mistral NeMo with it in our default models for answering new questions [3].
[1] https://pvq.app/leaderboard
[2] https://openrouter.ai/microsoft/phi-4
[3] https://pvq.app/questions/ask
Edit: they have a blog post https://pvq.app/posts/individual-voting-comparison although it could go deeper
We would have liked to pick a neutral model like Gemini which was fast, reliable and low cost, unfortunately it gave too many poor answers good grades [1]. If we had to pick a new grading model now, hopefully the much improved Gemini Flash 2.0 might yield better results.
[1] https://pvq.app/posts/individual-voting-comparison#gemini-pr...
The one red-flag w/ Phi-4 is that it's IFEval score is relatively low. IFEval has specific types of constraints (forbidden words, capitalization, etc) it tests for [2] but its one area especially worth keeping an eye out for those testing Phi-4 for themselves...
[1] https://docs.google.com/spreadsheets/u/3/d/18n--cIaVt49kOh-G...
[2] https://github.com/google-research/google-research/blob/mast...
Did it have a different license before? If so, why did they change it?
https://ollama.com/vanilj/Phi-4
Then a quick search revealed you can as of a free weeks ago
https://ollama.com/blog/structured-outputs
Lots of other models will work nearly as well though if you just give them a clear schema to follow and ask them to output json only, then parse it yourself. Like I've been using gemma2:9b to analyze text and output a json structure and it's nearly 100% reliable despite it being a tiny model and not supporting tools or structured output officially.
Edit: so for example of you want the unsloth "debugged" version of Phi4, you would run:
`$ollama pull hf.co/unsloth/phi-4-GGUF:Q8_0`
(check on the right side of the hf.co/unsloth/phi-4-GGUF page for the available quants)
For the Phi-4 uploaded to Ollama, the hyperparameters were set to avoid the error. The error should stop occurring in the next version of Ollama [2] for imported GGUF files as well
In retrospect, a new architecture name should probably have been used entirely, instead of re-using "phi3".
[1] https://arxiv.org/html/2412.08905v1
[2] https://github.com/ollama/ollama/releases/tag/v0.5.5
[1]: https://news.ycombinator.com/item?id=42660335 Phi-4 Bug Fixes
Also on hugging face https://huggingface.co/microsoft/phi-4
Deleted Comment