`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%.
And agreed, people expect $ they invest into computer systems to do much better than their bad & avg employees. AI systems get the added challenge where they must do ~100% on what non-AI rules would catch ("why are you using AI?") + extra lift from AI ("what did this add?"). We generally get evaluated on matching experts (low bar), and exceeding them (high bar). Comparing to average staff is, frustratingly, a breakout.
Each scenario is different obviously..
FWIW – across all of these, we already do automated prompt rewriting, self-reflection, verification, and a suite of other things that help maximize reliability / quality, but those tokens add up quickly and being able to dynamically switch over to a smaller model without degrading performance improves margin substantially.
Fine-tuning is a non-starter for a number of reasons, but that's a much longer post.
OpenAI does great at training for general tasks, and we should not be disappointed when specialized tasks fail. Interestingly, openai advertises increasingly many subjects they are special casing like math, code, & law, and so holding them to standards is fair there IMO.
For specialized contexts openai doesn't eval on, these merit hiring consultants / product to add the last-mile LLM data & tuning for the specific task. And at least in my experience, people paying money for AI experts & tech expect expert-level performance to be met, and ultimately, exceeded..
1. Common themes in our development-time loop:
* We don't do synthetic data. We do real data or anonymized data. When we lack data, we go and get some. That may mean paying people, doing it ourselves, setting up simulation environments, etc.
* We start with synthetic judges, esp for scale tasks that are simple and thus considering smaller models like gpt-4o-mini (the topic here). Before we worry about expert agreement, we worry about gpt-4o agreement, and make evals that cover concerns like sample size and class imbalance...
* ... When the task is high value, e.g., tied closely to a paying customer deliverable or core product workflow, we invest more on expert evals, making calls like on how many experts and of what caliber. Informally, we've learned multiple of our teammates, despite good at what they do, can be lousy experts, while others are known for precision, even if not data people (ex: our field staff can be great!). Likewise, we hire subject matter experts as full-timers (ex: former europol/fbi equivs!), source as contractors, and, partner with our customers here.
* After a year+ of prompt engineering with different tasks, models, data, and prompt styles, there's a lot of rote tricks & standard practices we know. Most are 'static' -- you can audit a prompt for gotchas & top examples to fill in -- and a smaller number are like in the OP's suggestion of dynamic prompts where we include elements like RAG.
On the last point, it seems incredibly automatable, so I keep trying tools. I've found automatic prompt optimizers like dspy to be disappointing in being unable to match what our prompt engineers can do here: they did not do better then prompts we wrote as experts with bare bones iteration, and leaning into the tools failed to get noticeable lift. I don't think this is inherent, just they're probably eval'ing against people we would consider trainees. Ex: I see what stanford medical fellows+phds are doing for their genai publications, and they would probably benefit from dspy if it was easier, but again, we would classify them as 'interns' wrt the quality of prompt engineering I see them doing behind-the-scenes. I'm optimistic that by 2026, tools here will be useful for skilled AI engineers too, just they're not there yet.
2. It's a lot more murky when we get into online+active learning loops for LLMs & agentic pipelines.
E.g., louie.ai works with live operational databases, where there is a lot wrt people + systems you can learn from, and issues like databases changing, differences in role & expertise, data privacy, adverserial data, and even the workflows change. Another area we deal with is data streams where the physical realities they're working with changes (questions+answers about logs, news, social, etc).
IMO these are a lot harder and one of the areas a lot of our 2025 energy is going. Conversely, 'automatic prompt engineering' seems like something PhDs can make big strides in a vacuum...
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:
- 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".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.
Human beings have agency, and we are very good at rolling with the punches. We've survived waves of automation for hundreds of years. I'm much more confident that we will continue to find ways to use these things as tools that elevate us, not replace us.
I really hope the dystopian version doesn't come to pass!
The difference with past technological breakthroughs is that they augmented what humans could do, but didn't have the potential to replace human labor altogether as AI does. They were disruptive, but humans were able to adapt to new career paths as they became available. Lamplighters were replaced by electrical lighting, but that created jobs for electrical engineers. Carriage drivers were replaced by car drivers; human computers by programmers, and so on.
The invention of AI is a tipping point for technology. The only jobs where human labor will be valued over machine labor are those that machines are not good at yet, and those where human creativity is a key component. And the doors are quickly closing on both of those as well.
Realistically, the only jobs that will have some form of longevity (~20 years?) are those of humans that build and program AI machines. But eventually even those will be better accomplished by other AI machines.
So, I'm really curious why you see AI as the same kind of technology we've invented before, and why you're so confident that humanity will be able to overcome the key existential problems AI introduces, which we haven't even begun to address. I don't see myself as a pessimist, but can't help noticing that we're careening towards a future we're not prepared to handle.
The cost/benefit for the labor/middle/low classes is at best low right now. I define that as someone who needs to trade time to continue surviving as an ongoing concern even if they have some wealth behind them.
I think the outcome where any form of meritocratic society gives way to old fashioned resource acquisition based societies is definitely one believable outcome. Warfare, land and resource ownership - the old will become the new again.
But my main argument is against the notion that this technology is the same as the ones that came before it, and that it will undoubtedly lead to a net better future. I think that is far from certain, and the way things are developing only leads me to believe that we're not ready for what we're building.
That being said, I can't see any world where there isn't mass ontological shock/hysteria, mass unemployment, and unrest at least for a few years, and I feel like it is definitely the kind of event you take active measures and preparations for beforehand.
And so do I, but like the golden rule of camping you should prepare for the worst and hope for the best!
I think there’s a less than 5% chance that this goes well, and that’s only if we get a series of things to go extremely well. And frankly, we’re tracking along the extremely bad path so far.
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
For coding: qwen2.5-coder:14b (9G file)
Misc. experiments, runs fast: llama3.2:latest )2 G file)
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.
Whether it be Facebook, Instagram, Threads, Messenger, WhatsApp, etc. their focus is to acquire users, keep them in their platforms, and own their content - because /human attention is fundamentally valuable/.
Meta owns 40% of the most popular social media platforms today, but their attention economies face great threats: YouTube, TikTok, Telegram, WeChat, and many more threaten to unseat them every year.
Most importantly, the quality of content on these platforms greatly influences their popularity. If Meta can accelerate AI development in all forms, then it means the content quality across all apps/platforms can be equalized - video on YouTube or TikTok will be no more high quality than on Facebook or Instagram. Messages on Threads will be no more engaging than that on Twitter. Their recent experiments with AI generated profiles[0] signals this is the case.
Once content quality - and luring creators to your platform - are neutralized as business challenges that affect end users lurking on the platform and how effectively they can be retained, then it becomes easier for Meta to retain any user that enters their platforms and gain an effective attention monopoly without needing to continue to buy apps that could otherwise succeed theirs.
And so, it is in their benefit to give away their models 'for free', 'speed up' the industry's development efforts in general, de-risk other companies surpassing their efforts, etc.
[0] https://thebaynet.com/meta-faces-backlash-over-ai-generated-...
Meta makes money from ads. To make more money, they either need to capture more of their users' time and show more ads, or show better ads that users click more often.Meta is betting on AI models making it easier to do both.
Better generative AI means you can make more ads faster, which means there are more ad variants to a/b test across, which means it's easier to find an ad that users will click.
To make users stay on their platforms, Meta figures out what content will keep them there, and then shows them that content. Before gen AI, they were only able to show existing content from real users, but sometimes the "ideal" thing for you hasn't been created yet. They bet on the fact that they'll be able to use AI to create hyper-personalized content for their users that engages them better than human-made content.
It might be that many people’s aesthetic sensibility is that AI-generated content is slop, but I’d still bet that tailored-perfectly-to-you content (and ads) will be highly engaging
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.
> 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?
I don't really think they do, because to me it seemed pretty much since GPT-1, that having callbacks to run python and query google, having "inner dialog" before summarizing an answer and a dozen more simple improvements like this are quite obvious things to do, that nobody just actually implemented (yet). And if some of them are not obvious per se, they are pretty obvious in the hindsight. But, yeah, it's debatable.
I must admit though, that I doubt that this obvious weakness is not obvious to the stakeholders. I have no idea what the plan is, maybe what they gonna have that Anthropic doesn't is gonna be a nuclear reactor. Like, honestly, all we are pretending to be forward-thinking analysts here, but in reality I couldn't figure out that Musk's "investment" into Twitter is literally politics at the time of it happening. Even though I was sure there is some plan, I couldn't say what it is, and I don't remember anybody in these threads expressing clearly enough what is quite obvious in the hindsight. Neither did all these people like Matt Levine, who are actually paid for their shitposting: I mostly remember them making fun of Musk "doing stupid stuff and finding out" and calling it a "toy".
What's the distinction? What kind of functionality do they offer that other models don't?
Lots of products have been successful without a technical moat. Facebook has network effects, Apple has UX (though silicon has become a technical advantage if not moat), Adobe has “everyone knows how to use these tools” switching costs, Google has brand synonymous with search.
Companies are betting that models will be commodities but AI products will be sticky.
I could switch to a different provider if I needed to maybe with cheaper pricing or better models but that doesn't mean OpenAI doesn't offer a "product".
Their unique value-adds are the Chat GPT brand, being the "default destination" when people want AI, as well as all the "extra features" they add on top of raw LLMs, like the ability to do internet searches, recall facts about you from previous conversations, present data in a nice, interactive way by writing a react app, call down to Python or Wolfram Alpha for arithmetic etc.
I wouldn't be surprised if they eventually stop developing their own models and start using the best ones available at any given time.
The "consumer conversational AI space" only exists right now as a novelty, not a long-term market segment. In the not too distant future that space will be covered for most users for free by their hardware manufacturers, and the number of people willing to pay a monthly subscription to a third party will drop even further than it already has.
Default destination for many is still just Google, and they've added AI to their searches. AI chat boxes are shoehorned into a ton of applications and at the end of the day it'll go to the most accessible one for people. This is why AI in Windows or in your Web Browser or on your phone is a huge goal.
As far as extra features, chat GPT is a good default, but they're severely lacking compared to most other solutions out there.
That is the reason they are making products so that people stay on the platform.
This means that in a world where AWS/Azure/GCP all compete in the compute and the models themselves are commodities, AI isn't a product, it's a feature of every product. In that world, what is OpenAI doing besides being an unnecessary middleman to Azure?
I'd agree there isn't much money in it. OpenAI should probably milk the revenue they get now and make hay while the sun is shining. But their apparent strategy is to bet it all on finding another breakthrough similar to the switch from text completion to a chat interface
OpenAI isn't losing yet because their models are still marginally better and they have a lot of inertia, but their API isn't going to save them.
> But their apparent strategy is to bet it all on finding another breakthrough similar to the switch from text completion to a chat interface
I'm still convinced that their strategy is to find an exit ASAP and let Altman cash out. He's playing up AGI because it's the only possible way that "AI" becomes a product in its own right so investors need to hear that that's the goal, but I think he knows full well it's not in reach and he can only keep the con going so long. An exit is the most profitable way out for him.
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
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 only judges that matter at this stage are humans. Maybe someday when we have models that humans agree are reliably good you could use them to judge lesser-but-cheaper models.
It should be noted that Mixtral 8x7B didn't grade its own model very high at 11th, it's standout was grading Microsoft's WizardLM2 model pretty high at #2. Although it's not entirely without merit as at the time of release it was Microsoft's most advanced model and the best opensource LLM available [1]. Which we also found generated great high quality answers which I'm surprised it's not more used as it's only OpenRouter's 15th most used model this month [2], although it's received very little marketing behind it, essentially just an announcement blog post.
Whilst nothing is perfect we're happy with the Grading system as it's still able to identify good answers from bad ones, good models from bad ones and which topics models perform poorly on. Some of the grades are surprising since we have prejudices on where models should rank before the results are concluded, which is also why it's important to have multiple independent benchmarks, especially benchmarks that LLMs aren't optimized for as I've often been disappointed by how some models perform in practice vs how well they perform in benchmarks.
Either way you can inspect the different answers from the different models yourself by paging through the popular questions [3]:
[1] https://wizardlm.github.io/WizardLM2/
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?
Then a quick search revealed you can as of a free weeks ago
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://news.ycombinator.com/item?id=42660335 Phi-4 Bug Fixes
Also on hugging face https://huggingface.co/microsoft/phi-4
For context: I've made some simple neural nets with backprop. I read [1].
The brief of it is by curating a smaller synthetic dataset of high quality from textbooks, problem sets, etc. instead of dumping a massive dataset with tons of information.
Does this mean the model was trained without copyright infringements?
llama.cpp basically dropped support for multimodal visual models. ollama still does support them, but only a handful. Also ollama still does not support vulkan eventhough llama.cpp had vulkan support for a long long time now.
This has been very sad to watch. I'm more and more convinced that vllm is the way to go, not ollama.
Only having one model host (hugging face) is bad for obvious reasons (and good in others, yes, but still)
Ollama offering an alternative as a model host seems quite reasonable and quite well implemented.
The frontend really is nothing; it’s just llama.cpp in a go wrapper. It has no value and it’s not really interesting, it’s simple stable technology that is perfectly fine to rely on and be totally unexcited or interested in, technically.
…but, they do a lot more than that; and I think it’s a little unfair to imply that trivial piece of their stack is all they do.
Whilst it's now a UX friendly front-end for llama.cpp, it's also working on adding support for other backends like MLX [1].
If you care about running efficiently on your hardware, then llama.cpp is they way to go, not ollama.
Now, applications like ollama obviously need to exist, as not everyone can run CLI utilities, let alone clone a git repo and compile themselves. Easy to use GUIs are essential for the adoption of new tech (much like how there are many apps that wrap ffmpeg and are mostly UI).
However, if ollama are mostly doing commodity GUI things over a fully fleshed-out, _unique_ codebase to which their very existence is owed, they should do everything in their power to point that out. I'm sure they're legally within their rights because of the licensing, but just from an ethical perspective.
I think there is a lot of ill-will towards ollama in some hard-core OG LLM communities because ollama appears to be attempting to capture the value that ggerganov has provided to the world in this tool without adequate attribution (although there is a small footnote, iirc). Basically, the debt that ollama owes to llama.cpp is so immense that they need to do a much better job recognizing it imo.
I use them because they run as a systemd service with a convenient HTTP API. That's been extremely helpful for switching between GUIs.
I also like their model orgazation scheme, and the modelfile paradigm. It's also really handy that it loads and unloads models when called, which is handy for experimentation and some complex workflows eg embedding followed by inference.
Is llama.cpp doing 100% of the "heavy lifting"? Sure, but some light lifting is also needed to lower the activation threshold and bring the value to life.
I would not use llama.cpp, it's simply too cumbersome.
If Ollama did not exist, I would have to invent it.
Is it not "innovative"? Who cares! I want it. Commodity GUI? Again, I don't think they have a GUI at all. Are you maybe thinking of OpenWebUI?
Now, I don't use AI that much, I could totally live without this. But if it weren't for the robust one-liner I probably wouldn't use local LLMs at all.
With llama.cpp or llamafile, I was constantly having to look up a model's paper, documentation or other pages to see what the recommended parameters were, recommended templates were, and so on. My understanding is that GGUFs were supposed to solve that, yet still I was getting poor results.
You know, I don't know all the details or if there's any difference between what Modelfiles are for versus what GGUF metadata is for, but my experiences with Ollama have been that it just worked. It took me a while to even try Ollama, because the expectation is that it would simply be another interface on top of the same issues.
There are things I don't like about Ollama, but mostly they were easy to work around by writing a few scripts. Not using any web UI with it at all.