OpenAI’s Second Day of Its 12 Days of Innovation Focuses on Fine Tuning Models for Expert Tasks, Google CEO Predicts Profound Changes to Search in 2025 and Mozilla Rebrands As It sStruggles For Relevance
Welcome to Hashtag Trending, I’m your host, Jim Love. Let’s get into it.
OpenAI’s second day of its 12 day innovation announcement focused on a new feature that is in Alpha testing and will be released this next year, called Reinforcement Fine Tuning. FIne tuning is not a new concept, after a model runs through its training sequence which involves ingesting a huge amount of data, the models are often fine tuned.
In fact, OpenAI used a form of reinforcement learning with human feedback (RLHF) to develop and fine-tune earlier versions of its models. RLHF used human evaluators to guide the model’s behavior toward desirable outcomes.
But this new addition, Reinforcement Fine Tuning is another level of model optimization that can be used take the base frontier model and tune it in a way that gives it additional expert abilities in a particular domain.
A form of this training has already been used at OpenAI, and as the presentation outlined, that’s how they’d take their existing models from performing at high school level to a PhD level of expertise.
The differentiation here is twofold:
First, OpenAI is planning on making this available more widely, eventually allow anyone to use this to create their own expert model.
Second, and most amazing, is that the number of examples that are needed. It can be as little as a dozen or even a few dozen examples that are needed.
So for example, OpenAI is working with Thomson Reuters and their unique dataset of legal knowledge to create training for what they are calling a co-counsel AI – an AI that could do some complex legal analysis.
Now, on our story last week we mentioned that Day 1 focused on areas where OpenAI is perceived to have a competitive disadvantage – coding being one of them where ChatGPT is largely regarded as inferior to Anthropics ClaudeAI.
It’s interesting that this second day started with the example of a legal AI expert. Why? Well Elon Musk has been trumpeting that is X AI is being trained to be an expert legal assistant. So this takes Musk on head first.
And there is no love lost between Musk and Sam Altman at this point. In fact at the same time that Musk was announcing the specialty in legal expertise, Musk also made a deal with Nvidia to buy a billion dollars worth of GPUs, which is going to suck up a lot of the market and potentially create problems for OpenAI. One industry watcher reported that Altman had questioned how Musk had essentially pulled one over on Microsoft, who OpenAI partners with.
So minor digression, but interesting to watch how OpenAI is using these announcements.
But – and there is a qualification here – a demo, as we all know, is not evidence of real word ability. But after they had done the mention of the legal partnership they pulled a Steve Jobs moment and did that “one more thing” where the crew of engineers, without Sam Altman this time, brought in a professor from Berkely to demonstrate how important this new facility would be – and they carefully took it from the commercial/legal area to a much more altruistic example.
Their use case was how this could be applied to rare genetic diseases, which the researcher was clear may be rare in terms of a relatively few cases of each individual disease, but that, in aggregate affect over 300 million people world-wide, which is almost the population of the United States.
They demonstrated how the model could be trained to increase its expertise in diagnosis. They took disease information from a number of publications and curated it.
Then they applied the Reinforcement Fine Tuning and demonstrated the increased ability of the model. They were quite clear that the results were not just that the model now could recognize new symptoms, — it could actually generalize this knowledge to new circumstances. So you could present symptoms and it could score highly on situations it had not seen in the original training data.
How effective were these. The demo did show a marked improvement. It correctly diagnosed the genes that were associated with a set of symptoms not in its training.
They took the o1 mini model which is quite small and after the training, it outperformed the 4o1 production model, which is their very best model. That’s quite a feat.
Now to reduce the hype on this, while the expert they used did say that this was an improvement, he was careful to say that it might augment, but not replace existing tools used to diagnose these ailments and match them with the genes that cause them.
While the team of AI engineers were excited, their outside expert was interested, but hardly “over the moon.” That could be his personality, or it could be that this is still and early step.
He was complimentary about the models ability to reason and deal with problems it hadn’t seen. He was also impressed with the fact that the model also showed its reasoning.
Again, this is a very early announcement but OpenAI is inviting other researchers and organizations with great potential use cases to participate and work with them over the next few months to develop this further and create expert models.
What wasn’t clear to me was whether this training would be used to develop what we assume would be proprietary expertise on top of the base model. We presume that Thomson Reuters is a business partner and isn’t giving up their expertise for free, but I was left wondering if this would all be specialize proprietary additions or if some of this additional training would be used to advance the overall model.
We’ll find out, I’m sure.
That’s the big story. For what little other oxygen there was for other stories,
Google CEO Predicts Profound Changes to Search in 2025
Google CEO Sundar Pichai took a poke at Microsoft and OpenAI when he announced that the company’s search engine will undergo significant changes in 2025, tackling more complex questions than ever before. Speaking at the New York Times DealBook Summit, Pichai said, “You’ll be surprised, even early in ’25, at what Search will be able to do compared to today.”
This follows Google’s ongoing AI-driven overhaul of its search engine, which now includes AI-generated summaries and features like enhanced video search via Google Lens. The company is also preparing to roll out updates to its Gemini AI model, designed to compete with Microsoft and OpenAI.
Pichai addressed Microsoft’s AI advancements, particularly its partnership with OpenAI, saying, “I’d love to do a side-by-side comparison of Microsoft’s own models and ours.” Which Pichai went on to emphasize, Microsoft was using “somebody else’s model.” Oooh. Burn. He emphasized Google’s commitment to remaining at the forefront of AI innovation.
As competition in AI search intensifies, 2025 could be the most competitive year for search we’ve ever seen as Google tries to catch up with or exceed Perplexity.ai and OpenAI’s search.
Hey, maybe competition really is a good thing.
Mozilla struggles for attention and even relevance with new rebranding
Mozilla’s recent rebranding aims to emphasize its commitment to an open and accessible internet, highlighting its role beyond the Firefox browser. The new brand identity, developed with Jones Knowles Ritchie, features a refreshed logo and custom typefaces, reflecting Mozilla’s activist spirit and mission to “Reclaim the Internet.”
However, the timing of this rebrand coincides with significant organizational challenges. In November 2024, the Mozilla Foundation laid off 30% of its workforce, eliminating its advocacy and global programs divisions. This restructuring raises concerns about Mozilla’s capacity to fulfill its mission of promoting a free and open web.
Financially, Mozilla has faced hurdles, with a substantial portion of its revenue—86% as of late 2022—coming from a search partnership with Google. This dependency is precarious, especially amid legal challenges to such agreements. citeturn0news23 Additionally, Firefox’s market share has declined, dropping to 2.65% in 2024 from nearly a third of users in 2009. citeturn0news23
While the rebranding seeks to revitalize Mozilla’s image, the organization must address these financial and operational challenges to maintain its relevance and effectively advocate for internet openness.
For those of us who have been Firefox fans in the past but now find ourselves on Chrome more times than not, it’s going to be a tough battle for Mozilla.
And that’s our show for today.
Reach me at editorial@technewsday.ca
I’m your host Jim Love, have a marvelous Monday.
