Conor Dragomanovich的动态

It's been a few days since OpenAI released their new o1 set of models. It's not called 'GPT5', so is it actually a big change? YES!! Here's my quick(ish) take. Most of the models so far have been using very similar methodologies and training data. If you track the general performance of these models on benchmarks, they are all reaching about the same 'GPT-4' level of capability. Meta's Llama, Anthropics's Claude, Google's Gemini ... all broadly similar. OpenAI's o1 is something new and the befuddlement around how they've exactly done it means that the architecture and approach is a bit different too. The way I think about this is related to human 'System 1' and 'System 2' thinking as popularised in the book 'Thinking Fast & Slow'. Basically, our system 1 thinking uses shortcuts and estimates to be super-fast whereas System 2 is more thoughtful and methodical. System 1 is great at lots of things - quickly spotting possible danger in the shadows for instance - but it also can make a bunch of errors, and I line this up with the idea of current LLMs suffering from 'hallucinations'. Here's my favourite example of a System 1 hallucination. Read this puzzle quickly and grab the first answer that's in your head: "A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?" Most people immediately say that it's obviously 10c ... but, you're wrong! Our system 1 thinking actually simplifies this question for the sake of speed and gets it wrong. The answer is actually 5c - write it down and you'll see. So, OpenAI's model is the first AI model I've seen that specifically tackles the idea of System 2 thinking. It doesn't simply spit out the most obvious or probable answer. It thinks, it checks again, and thinks again, until it's pretty sure that it's covered all the bases and hasn't jumped to conclusions. So, this is actually a very meaningful shift and important step towards more useful and reliable AI. As always, I'll wrap up with an encouragement to everyone to get across this stuff, to try it out, and to get skilled up. Our world will change dramatically over the next decade or so and it's started already. All the best. https://openai.com/o1/

New OpenAI o1—a new series of AI models have been released about and hour ago. The model shows the thinking process, how to answer and how it answered. Which is helpful for prompting. Confession: have few paying subscriptions for OpenAI in a proper geek way. With AI models it is give and take. The more people interact, models can learn from interactions. What i give is far less than what I get from OpenAI so i like for them to be in business and keep going until loosing my work because the AI models can do all the work one day :-) / :-(

New models come out literally weekly at this point, not just LLMs but also embedding models (the work horse of RAG applications)! Make sure when you're building your infrastructure you're building it in a way that makes it trivial to swap between these models so you can be sure you're delivering the most value in your AI adoption. If you are interested in what this means in practice, my dms are always open to discuss AI infrastructure and enterprise adoption

At Explain, we’ve designed our product to welcome LLM improvements enthusiastically, rather than fear they'd render it obsolete. From day one, we've built it to be model and provider-agnostic. Our approach was to use our own benchmark based on our use cases—summarization and RAG on complex documents—and to update the model we use with every major release. This strategy led us to switch from OpenAI to Mistral AI in June. We are now back to OpenAI and gpt-4o-mini. This is because of one of the key trend of the last 0,5 year (see this great post by Philipp Schmid from a few weeks ago https://lnkd.in/e2da2kSw): for a given performance level, models are getting faster and cheaper. When benchmarking models on our benchmark, gpt-4o-mini is not significantly more accurate, but is x2.1 faster and twice as cheap, so it was a no brainer as latency is key for us (as for many real life use cases). Several recent papers (see in comments) show convincingly that fine-tuning can achieve a lot to enhance LLM performances. I still think it is the most promising direction in R&D in the medium/long term. However, for LLM-based products I would point to the fact that there are many low-hanging fruits to be gained by simply choosing the right model at the right time.

OpenAI released a new GPT-4o snapshot with another price cut. I've been saying that this is a race to zero and AI will be free for a while, and I love collecting a new datapoint to support my thesis every few months. If you're thinking about an application that is too expensive to run with today's AI prices, you can make a business case that it will cost 10x less within a couple of years max. A few interesting questions to ask: - What are the margins for model providers today? - Why is OpenAI slashing costs from a market leader position?

As the ecosystem develops, models will be seemingly fit for purpose for specific business and technological use cases. Maintaining flexibility to incorporate and train the most appropriate model for your developers to develop enterprise applications will be of strategic importance, or we will repeat the silos and "modernization" needs of the past.

RAG lives and dies on the quality of it's retrieval. Snowflake Arctic's embedding models are a set of highly efficient retrieval models that prioritise quality at low cost to run - making high quality search available for all Enterprise use cases. Oh, and on the subject of *all*, we've open sourced the entire set. ?? The web search expertise of the Neeva team have played an enormous part in this story - I highly recommend reading this blog on how these models were trained end-to-end on Snowflake: https://lnkd.in/gzwtDkvJ

Tool use or function calling really changes how we think of the capability AI models! OpenAI had started this a year ago with function calling, but it was brittle and didn't really work. Now Anthropic's Sonnet 3.5 model has really taken it to the next level. Kudos to them! Now, the entire outside world and interaction models are accessible to AI models. This allows incredibly more powerful agents and chains -- the AI model can really interact with the environment around them. Whether it's booking a restaurant res, or writing a file to your computer, or starting up a docker image -- it's all possible -- all managed by the AI model's reasoning.

I love seeing innovation in this space that is focused on problems that every company faces: getting the highest quality results *fast*, in an efficient way, and delivered in the easiest way possible! #DataCloud #OnePlatform

New OpenAI models dropped. Time to grab some lunch with a side of ??(not everyone will get this) Introducing OpenAI o1 We've developed a new series of AI models designed to spend more time thinking before they respond. Here is the latest news on o1 research, product and other updates. https://lnkd.in/eE4zbP2i