Jev appears to be a great new tool for developers writing AI native applications. It has the world knowledge and intelligence of a top tier LLM, but operates in a constrained way such that answers come back faster/more efficiently. Not for consumers, meant for developers.

OK, my last All-In Summit summary (see previous posts for more), took place yesterday but I’m just getting around to posting it now: Jensen Huang, Nvidia CEO. Standing ovation for Jensen. Everyone loves him. Was immediately asked about Dario’s essay. Jensen says it contains a false choice: Safety or creating frontier tech. You can do both at the same time. He also ridiculed the 10% prediction of humanity apocalypse saying it wasn’t grounded in science. It’s a made up number. He said saying things like that is very irresponsible. He gave a bunch of examples of wrong AI predictions: Radiologists out of work, 100% wrong. By now, 90% of code to be generated by AI, also wrong. 50% of entry job to be wiped out, wrong. Jensen said that frontier labs must take accountability for all these stupid predictions. Regulations should solve actual problems. Frontier labs are where the dangers come from since they have the most compute. So he called on them to just dig into details and get engineering & processes done better. RSI (recursive self improvement) is the newest boogieman. He doesn’t believe AIs will spiral out of control. In general, he gave a very grounded engineering perspective. Jensen talking up their acquisition of Hugging Face. He gave this statistic: $400B of VC money has gone into AI native companies last 6 months, 80% of which use open models. With respect to competing with China. He made the observation that the last Industrial Revolution in the 1700/1800s all came from European inventions, but Americans did a better job of exploiting them. The analogy is that even if China makes great open source models, and eventually leading process node chips, the US should be able to exploit AI better than China. Jensen definitely sees AI as a new Industrial Revolution. He said that Nvidia will go up as far they need to in the technology stack, but stay as low as possible. The rest of industry should do the AI end user and business applications. Will Nvidia develop frontier models? He said they already have five frontier models, but they are specialized. He cited Alpamayo for self driving cars and ESM-2 for protein folding. When specifically asked whether Nvidia chips could be manufactured at Tesla/SpaceX’s Terafab, he sidestepped and said you can’t discourage Elon from doing anything. He predicted that China will have advanced lithography by 2030. With regard to super intelligence, he stated that we’re already there if you look at narrow AI (eg. self driving cars and protein folding). Finally, on AI doomers, he commented that if you’re under 35 you’re likely to be a doomer, over 35, far less likely and he postulated that older people have just lived through so many “this time it’s different, the end of the world really is coming” busted narratives.

All-In Summit: Naveen Rao, CEO @unconvai Unconventional is one of half a dozen unicorns and startups that aim to completely replace conventional silicon, backprop algos, transformers, basically everything that powers our existing very powerful AI. They have a local connectionist training algorithm & hardware implementation (they showed off a picture and some specs from their just manufactured chip). Instead of encoding neuron weights in SRAM or HBM, it stores them in an analog dynamic environment of coupled oscillators. Meaning the compute and memory are part of the same substrate. What you get from this is the elimination of the Von Neumann bottleneck which costs a lot of energy. They have promised 1000x power efficiency over Nvidia performance within 2 years. The just manufactured chip is a proof of concept. No actual benchmarks were given (other than energy per image generation at an crappy image generation benchmark), so we don’t know its accuracy, chip density (both compute and memory), ability to train and inference LLMs, etc. So, some promise and advancement, but the jury is still very much out until we see some benchmarks. I tell people we’re still in the first inning with AI. Another analogy is that we’re still in Fortran era with goto statements. Our huge AI industry scaled way before we even started optimizing. Lots of disruptions are still expected, I just don’t know when and from where they’ll hit … yet.

More All-In Summit. Jensen was on stage talking when his assistant frantically waved at him to take a call. “Uh oh!” He said. It was Trump wanting to evangelize his own take on AI safety, which Jensen agreed with. They fumbled a bit getting Trump on the audience speakers and this was the rest of the call.

All-In-Summit 2nd speaker: Dina Powell McCormick president and vice chair Meta. Gave impassioned (and correct) defense of data center buildout showcasing how it has revitalized Richland Parish, Louisiana when they started their data center project there 20 months ago. Average salary there was $38K before the $20B data center build out project. Needless to say it has revitalized the economy. They just announced $50K bonuses for their teachers. They now have tons of inbound of the best teachers in the State wanting to work there. Louisiana is now courting data center builders from around the country to build in the State using Meta’s Richland Parish deal as the template. Water, of course, has been a nonissue. The data center uses less water than the farms that previously existed there used. Meta paid for its own electricity generation, storm mitigation upgrades, etc. Meta also touted their Workforce Academy training program for data center techs of all kinds. It’s been so successful that other companies like GM are wondering how they can get involved. JCal asks about Facebook lawsuit settlementment. 52 attorney generals agreement. Facebook now shuts off after 2 hours for children under 18. Wants YouTube and TikTok to join. Meta Glass, what’s the sticky use? Conversation focus is a big one. In a crowd, when you look at someone, the speakers will enhance that conversation and mute others. I could have used that yesterday on the shuttle bus when talking to an Aussie on a very noisy and crowded shuttle bus.

All-In Summit: First speaker, Satya Nadella, Microsoft CEO. On pausing AI: We should always endeavor to have humans in control. But recognize that we are growing intelligence that we really don’t understand fully, not building it. So you need competent devops to monitor/control it. Yes to third party testing. Wants broad diffusion of AI tech. Denigrates the 10% chance humanity gets wiped out, saying if you see a show stopper, just stop the show. So just develop responsibly. Already sees massive model capability overhang. Models are great already, but we haven’t made enough use of what we have. Still need to drive more business/consumer use. Wants KV cache standards so that you can share it among many different AI model families operating together. Frontier labs facing token compression price: applications will become much more viable economically.

Great post, spot on. There’s an actual historical precedent for this in AI. Waaay back in 1990, HNC was a tiny San Diego neural network software tools company. Yes, AI didn’t pop out of nowhere in 2022, it’s been around since the 1950s really. It’s just that hardware didn’t get powerful enough to interact with us in English until recently. But I digress. HNC realized that actually making an application using their AI was going to be far more profitable than pushing on a string trying to get clueless industry to use their advanced (for the time) AI. So they made a credit card fraud detection product using their primitive neural network tech of the time, and it worked great propelling the company/product into thousands of banks and credit card processors. Frontier AI companies have to make that same transition because whoever does it first will reap far more $ than the others. The play as an investor is to invest in the economy. Ie. The S&P 500 unless you do a lot of research to figure out the details.

The reason I'm a fan of @Ultraroboticsco is they do real world AI robotics NOW. This is live footage from a 3PL customer site with three Ultra bots coordinating to build boxes, pack, tape, label and place into the correct USPS, Fedex, etc. bins. Ultra writes all the back end software to integrate with the customer's order processing system, as well as the AI training unique to the application (building boxes, unique widgets, packing material, etc.). Customers are beyond thrilled to use Ultra. Not only is it hard to find employees who don't quit after a few months for these jobs, but when bought under a RaaS contract, payback time is ZERO. They save money immediately. Sound on.

New Tesla FSD technical talk. For the first time, we learn fascinating details of how they train their end to end AI model and it turns out it’s exactly how humans learn. When we first learn any complex task like driving, our conscious mind (cortex) is fully focused on all the little details our instructor has told us to focus on. Are you centered in your lane? Look out for stop signs. Pay attention to stop lights. We are so mentally overloaded that we don’t drive very well. But what we are doing through cortex guided practice is training our unconscious brain areas (basal ganglia and cerebellum) to drive without our conscious cortex having to pay attention to and direct everything. After many hours of practice we have effectively coded an end to end neural network in our brains that allow us to drive without much conscious thought at all. This is exactly how Tesla trains their end to end FSD v14. In the data center, they use labeled training videos which tells the model being trained to pay attention to lane markings, stop signs and street lights. But the resultant created model doesn’t have any explicit sections that look for these cues, it is all implicitly baked in the single end to end model that the car uses to drive. Before FSD 12, the AI running in the car did have a 100 different feature detectors, which is why it drove like a brand new driver. It turns out that none of that engineering work was wasted, they still use it, but only when training a new model from scratch. And yes, this has direct applicability to training Optimus. Much more info in the talk, including how VLAs fit into this picture. Start the video around 6m50s to skip stuff you already know.

@herbertong just interviewed me about The Boring Company, my favorite Elon private company (now that almost everything else is public!). If you're curious about what the big deal is with this company, give it a listen.

I listened to 85 minutes of The Economist’s interview of Elon so you don’t have to. Besides, it’s behind a paywall. Elon’s predictions: In five years, AI compute will exceed the sum of all human intelligence. In ten years, we will have reached the age of abundance. Money won’t matter. Everyone will have what they need or want (at least in economies that embrace AI). Ms. Beddoes tried to pin Elon down on how the economy will transform that way, but he wouldn’t get into specifics beyond noting that widespread AI robotics is a deflationary force. This means governments won’t need to raise taxes for universal basic income schemes, or, as Elon likes to call it, universal high income, since they will simply be able to print money to ward off deflation caused by the robot economy. She noted that Elon appears to have a more sanguine view of AI lately. He replied that he’s concluded superintelligent AI is now inevitable, so there’s no point trying to stop or slow it down, it can’t be done. We might as well enjoy the ride. The interviewer also noted that Mars no longer seems to be Elon’s overall ambition. He answered that his real mission was always to propagate and preserve human consciousness into the far future. Mars was just a vehicle for that. But now AI is a very important part of that goal. AI will necessarily be part of any future plan. And then came the oh-so-typical, increasingly tiresome part of most long journalist interviews: the interviewer constructs a straw-man version of Elon and argues against it. Elon carefully explained that he isn’t a raging far-right extremist, racist Nazi who kills puppies … and the journalist still didn't believe it. It is so effing tiresome. The lack of self-awareness on the part of journalists is off the charts. She complained about Elon’s supposed misperception of how dangerous London is, while remaining oblivious to the role she plays in creating the giant misperception of Elon as a person in her own writing. Elon defended his political views, saying he is for secure borders, locking up criminals, and balanced government spending, something even she had to admit didn’t sound crazy. And… that’s about it for an 85-minute interview. I couldn’t help but think that the next long-form interview Elon does should be conducted by an AI.

@herbertong posted his interview with Lars, Tesla head of engineering. Here are the key take aways that I heard, please let me know what I might have missed. “Cybercab is going to be scaled to levels that people haven’t realized yet.” They are currently doing infant mortality testing and validation on the Cybercab since it has such new components and manufacturing processes. Cybercab manufacturing line is 90% automated, factoring in everything. He expects Tesla to build more Cybercabs than any other vehicle they’ve built. Lars shoots down advertising yet again, he says Tesla products will speak for themselves. Advertising won’t change opinions, specifically talking about Robotaxi. Wouldn’t get roped into a discussion of selling the Cybercab this year (as opposed to using them in house for Robotaxi). Tesla is using AI a lot. They’ve made an internally trained agent on all the years of engineering knowledge about what works, what doesn’t, what went wrong, how do certain things etc. Basically an internal engineering chatbot with all of Tesla’s documented engineering knowledge. Also a similar one for their supply chain teams and one for their service centers to collect all the knowledge individual mechanics discover as they service the cars. They also use AI for anomaly detection in their manufacturing QA. AI detects subtle trends or confluence of events. They’ve even trained or are training an AI to listen for all the squeaks and rattles in the car as it comes off the assembly line. The car diagnoses itself (since it has internal microphones) as it heads to the final processing center. He called this “Full self hearing”. This allows Tesla to fix minor issues before customers get them. Optimus. The automated manufacturing lines had/have been constructed in Germany and the first line has landed and is/will be installed. Field acceptance testing of other lines is still happening in Germany. In theory they can get a line going in a week’s time from delivery from Germany. There are about 40 sub lines for Optimus. IMHO, this lines up with an August start of Optimus production. Fascinating tidbit from Lars. He said BOM cost for a car is higher than a robot. Now, I don’t know if he was talking in theory, in the future, or current BOM cost? If current BOM is less than a car already, that’s huge considering how capable Optimus is (it’s much more physically capable than a cheap Unitree for instance). Lars pushed back on bots being manufactured like a consumer device like a phone. He said it is a true 3d device, has safety risks, and is built much more like a car. He stated that Starlink is an obvious choice for Cybercab in the future for rural areas, but it would be a future enhancement. Finally, mark your calendar for Tuesday July 7 when Tesla will make some sort of announcement about future product/production scaling.

Recently four companies (@IneffableLabs , @amilabs, @thinkymachines, @unconvAI) raised over $4.5B to make AI chips several orders of magnitude better than NVIDIA … and they all raised the money with no clear roadmap of how they are going to accomplish this. For all their computation power, NVIDIA chips, TPUs, Trainium, Groq, all suffer from being massively power inefficient. They use way more power than they could if we had a different AI paradigm. The human brain is often cited as a counter example, running on a mere 20 watts of power. So these companies are all attacking this power problem from many different directions: New training algorithms using non-linear dynamics. Analog electronics instead of digital. Spiking neural nets. Compute in memory. Yann LeCun’s JEPA architecture. Multi-modal intelligence. Using Alpha Zero’s approach of not using any human training data to create a super learning algorithm. There’s no guarantee any of these companies will succeed. But if one of them does, it’ll massively disrupt the AI industry. It’s no surprise that NVIDIA and Google invested in some of these companies, they are hedging their entire revenue stream on the possibility of such disruption. Elon’s investment in Terafab, but more specifically on his R&D fab, is also a hedge against some breakthrough (the R&D fab is specifically going to investigate “new physics” for chipmaking). It won’t happen quickly even if a breakthrough happens tomorrow. It takes time to make new semiconductor devices, prove them out in a manufacturing lab, and ramp up production. We will get years of warning. The AI industry’s course is set for the next five years at least: massive data centers and a race to secure new massive energy infrastructure. But it’s nice to know there’s a lot of venture money working on something that's potentially massively more efficient.