Is AI Slowing Down?

Today on the AI Daily Brief, is AI actually possibly slowing down before that in the headlines? Salesforce hires 1,000 salespeople to sell AI salespeople. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. To join the conversation, follow the Discord link in our show. Notes. Welcome back to the AI Daily Brief Headlines edition. All the daily AI news you need in around five minutes. Today we are going to rip through a set of stories as the main episode is a little bit longer than usual. We kick it off with a story from Salesforce. Now Salesforce is pushing the agent story more than any other company out there. You might have heard CEO Marc Benioff absolutely positively screaming about the fact that the assistant era of AI has been a big letdown because of course it's really and was always really about agents all along. The news today is that Salesforce is hiring 1,000 people to push their new AI agent platform. Benioff said that the hiring surge is aimed at making the most of the amazing momentum of their new product. Adding AgentForce became available just two weeks ago and we're already hearing incredible feedback from our customers. One of the important pieces of context is that Salesforce has been downsizing their sales team over the past two years, so there is clearly a lot of excitement around this new product line. Of course, the joke everyone is making on Twitter is some variation of what Thomas Smollet wrote, Salesforce hires 1000 salespeople to sell their AI salesperson. Just writing that out sounds like a joke, but still, I think it's a really positive sign that there's so much demand that this is the actual direction they're headed. Meanwhile, over in Elon, World X is testing a free version of xai's Grok chatbot. Until now, the AI features of X have been only available to paying customers. However, over the weekend it seems that Grok was available to free users in select markets. TechCrunch confirmed that all of New Zealand had access to Grok, and it seems that Australia also has access to Now. Usage is Limited, with 30 text queries every two hours and three image analysis questions per day, but there's still a lot of excitement. Dr. Warface writes, If Grok becomes free, I will delete ChatGPT. Over in the land of Google, According to the Wall Street Journal, Google News exec Salish Prakash has resigned. The resignation comes as news publishers rethink their relationship with big tech. In the AI era, publishers have seen traffic and revenues decline, with many blaming AI for serving content summaries rather than driving clicks. The issue has become even more contentious with Google rolling out their AI overviews globally. Prakash was originally brought in two years ago from the Washington Post to bridge the gap between the two industries, and therefore many are wondering if his departure is a signal that the relationship is souring even further. Then again, there are clearly structural changes underway at Google following a year marked with losses in antitrust lawsuits, so it's hard to know exactly what's going on. One other interesting story out of Google the company's DeepMind division has unexpectedly released AlphaFold 3. The source code and model weights for the protein folding AI are now available for Academy use, which could further accelerate scientific progress and drug development. The surprise announcement comes weeks after the system's creators, Demis Hassabis and John Jumper, were awarded the Nobel Prize in Chemistry for their groundbreaking model. AlphaFold 3 seems like another paradigm shift for biological science. Version 2 could accurately predict protein structure based on nothing but their amino acid sequence. The new version can also model the complex interactions between proteins, DNA, RNA and small molecules. This should massively expand its usefulness in drug discovery and disease treatment. Traditional methods of studying these interactions often take months of lab work with no guarantee of success. This marks an important shift in computational biology. AI methods now outperform our best physics based models in understanding how molecules interact. Essentially, AlphaFold has gone from a specialized tool to a more general Swiss army knife for researching molecular biology. Over in the market realm, US companies are set to spend $30 billion on data centers as AI drives a construction boom. That amount has more than doubled since 2022 when ChatGPT was first released. Data center investment is now bigger than every other major infrastructure category, including hospitals, schools, hotels and transportation. It was a smaller category than each of these until late last year. Investment manager KKR believes the trend is exponential, expecting it to hit 250 billion per year globally within the next three to four years. Although stick around for our main story as there could be some implications on that front. That, however, is going to do it for today's headlines. Next up, the main Episode. Today's episode is brought to you by Plum. Wanna use AI to automate your work, but don't know where to start? Plum lets you create AI workflows by simply describing what you want. No coding or API keys required. Imagine typing out AI analyze my Zoom meetings and send me your insights in notion and watching it come to life before your eyes. Whether you're an operations leader, marketer or even a non technical founder, Plum gives you the power of AI without the technical hassle. Get instant access to top models like GPT4O, Cloudsonnet, 3.5, Assembly AI and many more. Don't let technology hold you back. Check out Use Plum. That's Plum with a B for early access to the future of workflow automation. Today's episode is brought to you by Vanta. Whether you're starting or scaling your company's security program, demonstrating top notch security practices and establishing trust is more important than ever. Vanta automates compliance for ISO 27001, SOC2GDPR and leading AI frameworks like ISO 42001 and NIST AI Risk Management Framework, saving you time and money while helping you build customer trust. Plus you can streamline security reviews by automating questionnaires and demonstrating your security posture with a customer facing trust center, all powered by Vanta AI. Over 8,000 global companies like LangChain, Leela AI and Factory AI use Vanta to demonstrate AI trust and prove security in real time. Learn more@vanta.comnlw that's vanta.comnlw Today's episode is brought to you by Superintelligent. Every single business workflow and function is being remade and reimagined with artificial intelligence. There is a huge challenge, however, of going from the potential of AI to actually capturing that value. And that gap is what Superintelligent is dedicated to filling. Superintelligent accelerates AI adoption and engagement to help teams actually use AI to increase productivity and drive business value. An interactive AI use case registry gives your company full visibility into how people are using artificial intelligence right now. Pair that with capabilities building content in the form of tutorials, learning paths and a use case library. And superintelligent helps people inside your company show how they're getting value out of AI while providing resources for people to put that inspiration into action. The next three teams that sign up with 100 or more seats are going to get free embedded consulting. That's a process by which our super intelligent team sits with your organization, figures out the specific use cases that matter most to you, and helps actually ensure support for adoption of those use cases to drive real value. Go to BeSuper AI to learn more about this AI enablement network. And now back to the show. Welcome back to the AI Daily Brief. Today we are having a really fascinating conversation that has implications for everything in AI from safety to the business of IT to technology frontiers to applicability to regulations, life and all sorts of interesting things bottled up in this one conversation around whether the pace of development of AI is actually slowing. Now, this is not a new conversation. It's something that has been debated for a very long time. And by the way, there are wildly divergent opinions here. There are some folks who think that you can just add more compute and more data and get more performant models. There are others who think that there is no path to AGI with the current architectures we're using. This debate is why people like Yann Lecun can, with a straight face, say that they believe that AI is dumber than a cat. And of course the implications of that view are significant. In the particular case of LeCun's point, it means that to him, all of these existential X risk type concerns are way, way overblown. Now, what no one is arguing is that AI isn't extremely valuable even in the state right now. One of the things that I've often said is that even if AI stopped developing in this moment, it would still take years for the world to adopt and adapt to the new processes that have been made available by just the technology that we have in this moment. However, it still would be fairly significant if in fact the rate of AI improvement was decreasing significantly. All right, so let's start by discussing where this conversation, or at least this iteration of the conversation, actually came from. The TLDR is that the information reported that OpenAI are apparently looking to develop new strategies to deal with the slowdown in AI improvement. Back in May, OpenAI CEO Sam Altman had told staff that he expected their latest Frontier model, which they're calling orion internally, not GPT5, he believed would be significantly better than last year's flagship model. At the time, OpenAI had only completed 20% of the training process for Orion, and it was apparently already on par with GPT4 level. However, now that there have been a few more months of work, things are starting to look a little bit different. According to the information sources, employees who have tested the Orion model found that although the performance does exceed the current models, the jump isn't nearly as profound as the improvement from, for example, GPT3 to GPT4. There are even some use cases where the model might not be consistently better than the previous state of the art. According to one employee, while Orion does better at language tasks, it doesn't necessarily outperform when it comes to coding, which, given how prominent coding is as a use case, could be a significant problem. Now what this is leading to is a different approach to thinking about scaling, the information writes in response to the recent challenge to training based scaling laws posed by slowing GPT improvements the industry appears to be shifting its efforts to improving models after their initial training, potentially yielding a different type of scaling law. Another approach, OpenAI has apparently created a foundations team to figure out how the company can improve their models given what is increasingly a scarce supply of novel training data. Those strategies reportedly include training Orion on synthetic data produced by other AI models, as well as doing more to improve the model's reasoning process during the post training process. One of the big challenges is that at this point the current frontier models have effectively been trained on pretty much all of the Internet's data. Still, by far the more interesting shift here is about thinking about scaling simply as a matter of pre training to instead thinking about the post training implications. And this gets back to something that we've talked about a lot here, the idea that O1, which is of course OpenAI's more advanced reasoning model, is a fork on the evolutionary tree of LLMs. Sam Altman when he announced it was clear that this reasoning model represented a separate path from GPT4. At the time it was assumed that OpenAI would parallel process these paths, but increasingly it seems like they're leaning into the reasoning path. And by the way, it's not just OpenAI who's dealing with this problem. A few weeks ago Alex Heath of the Verge reported, I've heard that the latest Gemini model isn't showing the performance gains that Google had hoped for, though I would still expect some interesting new capabilities. The chatter I'm hearing in AI circles is that this trend is happening across companies developing leading large models. Former OpenAI chief scientist Ilya Sutskever spoke to this issue, stating the 2010s were the age of scaling. Now we're back in the age of wonder and discovery. Once again, everyone is looking for the next thing. Scaling the right thing matters now more than ever. And this once again gets us to what Turing Award winner Yann Lecun was pointing out in last month's Wall Street Journal piece, the WSJ wrote. Lecun thinks that the problem with today's AI systems is how they are designed, not their scale. No matter how many GPUs tech giants cram into data centers around the world, today's AIs aren't going to get us to artificial general intelligence. There are also questions around whether synthetic data will actually create their own set of new problems, although that's far from clear as well. But overall, it still appears that OpenAI is planning to double down on their reasoning pathway. Last month at an appearance at TED AI in San Francisco, OpenAI researcher Noam Brown said it turned out that having a bot think for just 20 seconds in a hand of poker got the same boosting performance as scaling up the model by 100,000x and training it for 100,000 times longer. And while this does open up a new pathway for improvement, it also implies having a lot more computing power to serve answers. Sonia Huang, a partner at Sequoia Capital, said this shift will move us from a world of massive pre trading clusters towards inference clouds which are distributed cloud based servers for inference. Nvidia CEO Jensen Huang also referred to this pathway, stating last month, we've now discovered a second scaling law and this is the scaling law at a time of inference. So where has the discussion been? First of all, there has been some confirmation from others that this is bubbling behind the scenes. Jan Peleg wrote, heard a leak from one of the frontier labs, not OpenAI. They reached an unexpected huge wall of diminishing returns trying to brute force better results by training longer and using more and more data than what is published publicly. For others who have been saying this stuff for a while, they use this as a moment to reinforce their point. Professor Pedro Domingos wrote, scaling laws are S curves, not exponentials. Insead professor Jason Davis explained that a little further he said this may be the most important strategic question about AI Is the rate of improvement of LLM performance starting to decline? Are we seeing maturation in the S curve? The report from the Information gives us some data, but I think we really don't know and is at odds with many insiders working on foundation models, especially since the effective performance of LLM systems will depend on complementary innovations that extend its usefulness and perhaps many years of individuals and organizations finding the right use cases and configuring their workflows to leverage them. Others, including the Information founder Jessica Lesson, wondered what the implications are for the data center business. She wrote, if this continues, the slowing rate of gains in trading LLMs has huge implications. 1 thought I had this morning Will being at the bleeding edge in chips matter as much over time? Will China therefore catch up to the US more quickly? Others pointed out that of course there was always going to need to be something new. Psy on Twitter writes, pre transformation architecture AI had hit a brick wall. More insights and architectural tweaks will be needed to break successive barriers for improvement. It would be unwise to imagine that current LLM is the final solution. Perhaps the most salient message though, is just that we need to think about scaling and training in a slightly different way. Estreyantel on Twitter X says AI hitting a roadblock? Give me a break. As if the diffusion transformer breakthrough hadn't happened. As if O1 reasoning isn't groundbreaking on its own. The only thing reaching a plateau is pre training, and that's a milestone, not a ceiling. Pre training was never meant to be the holy grail of neural network optimization will inevitably move towards models that can actively learn for a reason. Haiderlowdeveloper writes. Even if current LLMs can't scale to AGI through compute increases alone, it's not a problem. OpenAI01 shows further innovation are possible and we haven't reached inference compute yet. Things are progressing even if LLMs don't magically scale to AGI with more compute. Professor Shelley Palmer writes, there's been a lot of chatter about the end of LLMs or LLMs starting to fail. Those are only headlines. Clickbait really? If you dig a bit deeper, you'll read that there are several schools of thought regarding how to efficiently scale the foundation models. If the goal is AGI, then just adding compute power to pre training may not be the best path to follow. Instead, researchers are exploring an alternative called inference scaling to achieve smarter AI inference. The process when an AI model generates outputs and answers can be optimized by having models, quote unquote think through multiple possibilities before settling on a response. This approach enables complex reasoning during real time use without increasing model size. OpenAI's recently launched O model is a good example. By enhancing inference, O can tackle tasks that demand layered decision making such as coding or problem solving in ways similar to human thought. Test time compute techniques make this possible, allowing models to dedicate more processing to challenging queries as needed. A move towards inference focused distributed cloud based servers instead of large centralized training clusters might create a more competitive chip landscape. While Nvidia is the go to chip maker for pre training hardware, there are a bunch of chip makers, amd, intel, et cetera that make hardware suitable for this new method of inference scaling. The key takeaway is simple. LLMs are not failing, they are evolving. Sensationalist headlines aside, this is how product development works. Every's Dan Schipper also pointed out that the framing of all of this was very problematic and that it didn't really reflect the conversations that he's having with people inside the labs, he writes. The message that this headline conveys is at odds with what people inside the big labs are actually feeling and saying. It's technically correct. But the takeaway for the casual reader that AI progress is slow is exactly the opposite of what I'm hearing. And indeed, ultimately there was so much chatter to that effect that the information actually ended up clarifying a little bit. Stephanie Palazzolo writes Seeing lots of discussion after our Scaling Laws article this weekend. To be clear, we're not saying the world is ending, just that researchers are having to find new ways to improve LLMs, I.e. test time, compute. That led to another article, Goodbye GPT, hello Reasoning. Oh, and this basically gets to the point that this isn't about a halt, it's about an alternative direction and a new path that shows more promise than the previous. Still, it's a super interesting conversation, one that we will continue to follow closely. For now though, that is the story. Appreciate you listening as always. And until next time, peace.