Was CES Really Just a Parade of AI Slop?
19:00 Share

Today on the AI Daily Brief, the good, the bad, and the ridiculous of AI at CES. And before that, in the headlines, Elon says that we've exhausted all AI training data. 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. We kick off today with a continuation of a discussion that really started in earnest at the end of last year. Elon Musk has weighed in, saying that he believes that AI models have run out of training data. Of course, there's been a growing opinion among AI experts that we've reached the saturation point of scaling pre-training data. Maybe most notably, last month, former OpenAI leader Ilya Sutskever said, we've reached peak data and there'll be no more.

The general premise is not that we've literally run out of data for models to ingest. It's that new troves of private data are likely to be duplicative of what's already contained in current training sets,

So it doesn't add much of anything to the model. In a live stream conversation on Wednesday, Elon Musk weighed in on the issue for the first time, saying, we've now exhausted basically the cumulative sum of human knowledge and AI training. That happened basically last year. Now, the solution to this isn't clear. Some experts think the industry needs to pursue new scaling models to improve AI. We've been talking a lot about reasoning mode or time test compute. Others believe there's merit in generating curated synthetic data based on the output of leading LLMs.

Musk believes that synthetic data could be a way forward, stating, "...the only way to supplement is with synthetic data. AI will sort of write an essay or come up with a thesis, and then will grade itself and go through this process of self-learning."

OpenAI hinted that they are already experimenting with this process during the launch of the O3 reasoning model. Their scientists quipped that they could instruct the model to improve itself, to which Sam Altman responded that maybe they shouldn't do that. There's also signs that synthetic data is already in heavy use. Gartner estimated that 60% of the data used for AI and analytics projects in 2024 was synthetically generated. Google, Anthropic, and Meta have all confirmed they are using synthetic data as part of their training. Microsoft provides a particularly interesting example with their Phi or Phi series of small models. The

Those models are trained using a mix of synthetic and real-world data that is carefully curated and labeled to create the minimum viable dataset. They're benchmarking off the charts relative to their size right now, which could suggest that training on synthetic data is a way of compressing models. Some labs are also using synthetic data as a cost-saving technique. AI startup Writer claims their Palmyra X004 model was trained entirely from synthetic data and cost around $700,000.

There's even been some suggestion that labs might start holding back their bleeding-edge models to prevent rivals from using them as a source of synthetic data. Musk did warn that the use of synthetic data comes with a risk. He noted that hallucinations are difficult to weed out, particularly as models start being trained on information too complex for human data labelers. Musk said it was, quote, "...challenging because how do you know if it hallucinated the answer or if it's a real answer?" He also discussed model collapse, the premise that using too much synthetic data risks the model degrading in quality.

Musk confirmed that, quote, when you start to feed a model synthetic stuff, you start to get diminishing returns. These tidbits of information could hint at some of the problems being encountered at XAI as they train the third generation of Grok. That model is likely to be the first in the world trained on a 100,000 GPU cluster and was intended to be released by the end of last year.

People are all over the place when it comes to this conversation. Investor Saad Hashimi writes, this has a mad gods outcome and I personally feel that my P doom has gone up. Saddle up, it'll be a bumpy ride. Evan Owen writes, synthetic data is the plastics of data. It's not part of the natural life cycle and we don't know yet what happens long term if we just keep recycling it into training sets. Very possible it creates hidden structures in the weights that manifest unexpectedly. Still, ultimately, it may be all that we have.

Grok, and that's G-R-O-Q Grok, head of sales Sunny Madra writes, synthetic data or bust? It will be an interesting conversation to watch. For now, let's jump over to Google, where that company has consolidated more of their AI team into DeepMind as they prepare for a pivotal year.

On Thursday, Logan Kilpatrick, the lead of product for Google's AI Studio, said that his team would be moving over to DeepMind, which is, of course, Google's AI research division. He posted, This move will allow us to double down on our already deep collaboration and accelerate the research to developer pipeline. The mission for our team stays the same. Build the world's best AI developer platform, which brings the latest models, tools, and techniques from Google to external developers so y'all can build the future. Google AI Studio is largely responsible for the Gemini API and productization of the company's AI technology.

In October, Google moved the team behind the Gemini app to DeepMind, stating, "...bringing the teams closer together will improve feedback loops, enable fast deployment of our new models in the Gemini app, make our post-training work proceed more efficiently, and build on our great product momentum." On this week's move, Gemini developer Janna Dogen remarked, "...our job will be making DeepMind's work publicly available in ways that wasn't possible before. While this is one of the biggest challenges in my life, it's potentially going to be the most rewarding."

When asked for specifics of what was coming, she commented, better APIs, more open source, more tools, you name it.

This, I think, follows what we've been hearing from Google, that the entire company is really focusing on AI as their most important priority. Alongside this announcement, we got the rollout of a feature called Daily Listen. It generates a five-minute podcast for users based on their personalized Discover News feed. Essentially, it's a one-click version of Notebook LM's generative podcast based on a user's interests. I would expect to see a lot of features like this over the next year from Google trying to bring popular AI features to a wider audience.

Lastly today, a slate of little announcements. Frontier Labs are getting off to a quick start in the new year with a full slate of feature updates and announcements this week. Microsoft has released their 5.4 model, fully open source with downloadable weights on Hugging Face. The small language model only has 14 billion parameters, but outperforms OpenAI's GPT-4.0 mini on benchmarks that test advanced reasoning and domain-specific capabilities. Of particular interest is the model's competitive performance as a coding assistant in a package small enough to run locally on consumer hardware.

OpenAI, meanwhile, have launched an easier way to customize the tone of ChatGPT's outputs. Users can now select from options like Chatty, Encouraging, and even something called Gen Z. You could already do this with any LLM by providing additional prompting, but ChatGPT now allows you to bake in the tone using a system prompt. Anthropic introduced a similar feature last month, and I actually think that these sort of UX changes are going to be significant in terms of how normalized the use of AI becomes.

Finally, XAI has begun testing a standalone iOS app for their Grok chatbot. Until now, Grok has been baked in as a feature on X. But as we've seen with OpenAI's advanced voice mode and vision capabilities for ChatGPT, having a standalone app allows for much more powerful use cases as an assistant. Grok's FAQ page is also teasing the launch of Unhinged Mode, which was first announced by Elon back in April.

The page says the model is intended to be objectionable, inappropriate, and offensive, much like the amateur stand-up comic who's still learning the craft. I'm sure that will be completely free of controversy, but for now, that is going to do it for today's AI Daily Brief Headlines edition. Next up, the main episode. Today's episode is brought to you by Vanta. Trust isn't just earned, it's demanded.

Whether you're a startup founder navigating your first audit or a seasoned security professional scaling your GRC program, proving your commitment to security has never been more critical or more complex. That's where Vanta comes in. Businesses use Vanta to establish trust by automating compliance needs across over 35 frameworks like SOC 2 and ISO 27001. Centralized security workflows complete questionnaires up to 5x faster and proactively manage vendor risk.

Vanta can help you start or scale up your security program by connecting you with auditors and experts to conduct your audit and set up your security program quickly. Plus, with automation and AI throughout the platform, Vanta gives you time back so you can focus on building your company. Join over 9,000 global companies like Atlassian, Quora, and Factory who use Vanta to manage risk and prove security in real time.

For a limited time, this audience gets $1,000 off Vanta at vanta.com slash NLW. That's V-A-N-T-A dot com slash NLW for $1,000 off.

If there is one thing that's clear about AI in 2025, it's that the agents are coming. Vertical agents by industry, horizontal agent platforms, agents per function. If you are running a large enterprise, you will be experimenting with agents next year. And given how new this is, all of us are going to be back in pilot mode.

That's why Superintelligent is offering a new product for the beginning of this year. It's an agent readiness and opportunity audit. Over the course of a couple quick weeks, we dig in with your team to understand what type of agents make sense for you to test, what type of infrastructure support you need to be ready, and to ultimately come away with a set of actionable recommendations that get you prepared to figure out how agents can transform your business.

If you are interested in the agent readiness and opportunity audit, reach out directly to me, nlw at bsuper.ai. Put the word agent in the subject line so I know what you're talking about. And let's have you be a leader in the most dynamic part of the AI market.

Welcome back to the AI Daily Brief. The Consumer Electronics Show is winding down in Las Vegas after a week crammed full of AI gizmos, announcements, and advancements. Now, as I mentioned earlier in the week, CES is kind of weird. It's open to the public, so it gets over 100,000 people going through the door. It's also catered to the consumer, so focuses on showcasing how tech is changing everyday products and even has direct sales on the show floor.

Mostly, though, it's absolutely gigantic with over 4,300 exhibitors this year. The question every year is whether CES is just going to be a look at whatever hype thing we'll not care about a couple years down the line or whether there's an actual there there with some really innovative new consumer products. What we're going to do today is first go through quickly a representative sample of all of the things that were on display. But then we're going to talk a little bit about what they all added up to.

At least one commentator had some fairly harsh words, so we're going to look at all that. Now, the most obvious theme, and why we're even talking about this here, was AI getting added to every appliance you could imagine. From AI fridges that can make their own shopping list, to TVs that can summarize the news or create a recipe from a cooking show, AI assistants were the hot new feature for appliances. Interestingly, some of the high-end AI features are starting to make their way into base-level products.

Barbecue tech startup Brisket unveiled a $400 wood pellet grill with wireless connectivity and an AI assistant. Brisket says their assistant can, quote, monitor, control, and automate the cook for you. Manual input is still required for things like wrapping the meat or spritzing it with water, but apparently the AI is smart enough to adjust the cook automatically if you miss the alert to avoid spoiling your barbecue. The price point is competitive right at the bottom of the range, which suggests that AI is quickly becoming a basic feature for every smart device.

Another big theme was AI assistance getting added to medical devices. The floor was filled with wearables getting an AI upgrade, but some companies are iterating on the idea. Withings unveiled a prototype smart mirror with a built-in health screening function that can test weight, heart, and lung health using a 360-degree scan of the user. Once it's finished, an AI-powered voice will explain the findings to you. Mashable's Christian DeLuper gave it a test run, and it correctly diagnosed a health issue he suffers from live on the show floor.

We'll come back to what was actually exciting about the show in a little bit, but this suggests we're starting to move to the point where health devices aren't just going to monitor our data, but actually analyze it to provide early warnings.

One of the big stars of the show was a new Roomba with a robotic arm. The idea is that the automated vacuum cleaner can also pick up your dirty socks or anything else weighing more than 300 grams. This one is a little more machine learning than AI, but it demonstrates how much the growth in machine intelligence is driving advances in consumer robotics. What's more, the device isn't a prototype. It's a fully functioning machine that will start shipping in February at a price point of $1,600.

There was a huge selection of AI-enabled smart glasses, headphones with built-in translation assistance, and other things that were firmly in the realm of science fiction just a few years ago, but are now part of fully functioning products.

It, of course, wouldn't be CES without a whole range of deeply strange products, though. A company called Spicer unveiled an AI-powered spice dispenser that can automatically season your meal based on the recipe and learning about your personal taste profile. A company called BirdFi, which advertises aggressively on social media, showcased an AI-enabled birdbath. In this case, the AI simply recognizes when a bird has arrived to cool off and snaps a candid picture.

A startup called Omni demonstrated their new AI wearable that competes with things like Rabbit and Friend. You can wear it as a pendant, but they're actually encouraging people to attach it to their temple for a sci-fi look. Omni claims the device can use a brain interface to understand when you're talking to it.

We will also skip over the seedier side of the conference, but suffice it to say, there are many NSFW things that AI is also going to be used for. So while AI was omnipresent at the conference, as you can see, some thought that many, if not most of the exhibitors, were simply slapping an AI label on an existing product. Most notably, Kyle Wiggers of TechCrunch declared that CES was full of in-real-life AI slop.

He wrote,

The next product on the list for criticism was an AI-enabled air fryer that can divine the correct setting by scanning a page from a recipe book. Summing up what he saw, Wiggers wrote,

He suggested that many of the products on display didn't represent what can be achieved by AI-enabled consumer tech, but rather what could be achieved with an extremely low R&D budget designed to get a product on the show floor quickly. I don't think it's a wrong critique or a bad critique, but I think there's a little more here. When we're looking at products, there's a difference between something that is actively bad, something that is simply not useful, and something which is useful enough to justify cost or a new device.

Some of this stuff is firmly in the camp of number three, not useful or overly complicated as to make it not useful. And I'm sure some is probably actively bad, but a lot of it is in that category of useful, but not necessarily switching cost useful.

To me, a more interesting question then is how likely it is that all of this stuff is just embedded in the future. In other words, are these use cases which feel somewhat banal even today, banal because they're just supposed to be a part or just likely to be a part of the core feature set of future devices?

With current user experience paradigms, a lot of these feel like features that would be left unused or solutions in search of problems. The theoretical increase in convenience isn't worth the cognitive load of shifting behaviors. However, I think things start to change when our main UX is just speaking out loud the things we want to happen. In other words, when we don't have to learn new features. It could be that a lot of this makes more sense.

There's also a generational dimension. It may be that in 20 years, the idea that we had to one by one select groceries on an app to be ordered to us or even crazier, actually go to a store will be just nuts. And that, of course, your fridge orders things for you. And all you have to do is have some preferences and a conversation at the beginning of each week about what you're feeling and interested in eating. It's also important to remember that we simply just have no idea how all of these things are going to play out.

During the IoT boom, no one really knew what the winning products would look like. No one really knew that the winning products would be stuff like smart home integrated light bulbs, garage doors, thermostats, and surveillance cameras. To take just one example, Ring was looking for $700,000 on $7 million on Shark Tank in 2013, and Kevin O'Leary tried to do it as a loan instead. Five years later in 2018, they sold to Amazon for a billion dollars. The point is, there is a lot to suggest that throwing spaghetti at the wall is the right thing to do at this stage.

What's more, I think it's important to not let the silly and stupid things at CES crowd out the genuinely important. Alongside the AI-powered spice dispenser, there were many products that gave a glimpse into the way that AI is going to change the world over the next few years. On top of that list has to be NVIDIA's AI supercomputer Project Digits. We covered this on Wednesday's show if you want a more complete rundown, but the idea is that the tiny supercomputer really stands out as something that could change the industry in very short order. The computer is about the size of a hardbound book. And

And coming in at $3,000, it's cheap enough to fit the budget of computer science students and AI enthusiasts. During his keynote speech, CEO Jensen Huang said the goal is to place an AI supercomputer, quote, on the desks of every data scientist, AI researcher, and student. NVIDIA claims the computer can run 200 B models locally and two networked together can easily handle the largest open source models currently available. This is a product then that is set to power a Cambrian explosion in the amount of tinkering, testing, and product development that can happen in the startup and academic setting.

This is already a use case that has demonstrated product market fit. At the moment, the most cost-effective way of building a local inference machine is strapping four Mac minis together. A single project digits machine has specs that seem three times as fast as that setup and can do it at less than half the cost.

Another point that's going somewhat under the radar is how fast consumer-level robotics is advancing. A Roomba with a robot arm stole the show largely because it's easy to see how that could be put to use immediately. But across the showroom floor, the robotic prototypes show that the sci-fi world is rapidly approaching. German Bionic unveiled a robo-exoskeleton capable of providing 80 pounds of lift assistance and a 20% boost to walking endurance. Other than making you look like a cyborg, these sort of devices could be a game changer for workers like baggage handlers and nurses who need to walk and lift all day long.

Pricing wasn't revealed, so this seems like something that's not quite ready to proliferate. And indeed, it seems like it was developed well before the recent boom in AI robotics. The company claims the device uses AI to adapt to its users' needs. However, they also mentioned that it was trained on thousands of hours of real-world usage. And that was just the tip of the iceberg in robotics that were on display there. Within a few years, the idea of having half a dozen robots whizzing through your house could be a completely normal thing.

Another of the big announcements from NVIDIA was their first series of open-source world models called Cosmos. Aside from enabling robots to be trained using simulations, the models have the potential to unlock a breakthrough in self-driving cars. Until now, the only way to train a self-driving algorithm has been recording and analyzing millions of hours of real-world driving data. That's why Tesla and Waymo have such a massive lead over legacy car makers who have barely gotten off the starting line. Jensen Huang said that the automotive simulation platform built around Cosmos promises to, quote, take thousands of drives and turn them into billions of miles.

Again, changing the scenario for how self-driving cars start to evolve. The point is that while, yes, a lot of this stuff may be the silly AI spice sprayer, there were many diamonds under the rough that do show, I think, where the world is headed. I think for most of us, it's worth trying to shelve a little bit of our cynicism. And I know that more than one of you is going to run out and get that Roomba with the robot arm. For now, that's going to do it for today's AI Daily Brief. Until next time, peace.