Friedberg explains LLMs are text predictors, reframing AI as prediction engines applied to genomics.
“They were able to suck up all the text on the internet and use that text to make a prediction on what the answer should be”
Friedberg explains LLMs are text predictors using internet data to generate word-by-word predictions, not true AI.
“They were able to suck up all the text on the internet and use that text to make a prediction on what the answer should be when someone asks a question word by word.”
Friedberg describes genome language models that can predict plant phenotypes from specific trait requirements like 90-day corn.
“I can now build models that some people are calling genome language models or, you know, whatever you wanna call them that can be predictive of a phenotype.”
Friedberg envisions artists defining boundaries for AI-generated media while preserving shared cultural experiences.
“the artist that creates the fundamental core of that cultural moment can define what is variable and what is not.”
Friedberg uses Arc Institute's EVO2 model in production at Ohalo to score plant DNA before phenotyping.
“We feed the whole DNA sequence of every plant in, and that model will tell us, Hey, these genes are dysfunctional or may have issues.”
Friedberg states new architecture reduces AI energy consumption by 70,000x compared to H100 chips.
“And the energy was reduced by 40,000 x compared to Jetson Nano, 90,000 x compared to RDX forty ninety, and a 70,000 x energy reduction for the same outcome over an h 100.”
Friedberg predicts new architecture enables sophisticated LLMs to run locally in robots without data center connections.
“So If this thing's even close to being correct, then you could run the most kind of sophisticated LLMs in a robot without it needing to be run out of a data center going forward and the robots can simply make a request for information from the Internet that they need.”
Friedberg reports new German AI architecture achieved 7000x speedup over Jetson Nano and 100x over H100.
“Their architecture led to a speed up of 7,000 x compared to the NVIDIA Jetson Nano, 300 x compared with NVIDIA RTX forty ninety, and then a 100 x compared to the NVIDIA h 100.”
Friedberg says ChatGPT's advanced voice has become unusable due to excessive politeness and data avoidance from overcaution.
“I used to use advanced voice on ChatGPT all the time. I cannot stand it anymore. I do not use it.”
Friedberg predicts SaaS will shift from per-seat to value-based pricing as software completes entire projects like services businesses.
“And I think it starts to look a lot more like a services type business where maybe the pricing is set up such that this thing will be completed for your business.”
Friedberg says AI researchers are surprised that making output recursive drives major model performance improvements.
“I saw a lot of computer scientists that have worked in AI for some time, I think, be a little bit surprised about this moment that we're in, that we're seeing such incredible strides in model performance just by making the output recursive.”
Friedberg predicts Google will launch virtual Cowork competitor within 90 days and sweep the market with G Suite integration.
“And once Google, like, has this integrated with G Suite and you have a virtual hosted version of Cowork, I think Google sweeps the market with the same competitor.”
Friedberg states Auto Research produced work equivalent to a seven-year PhD thesis that would be published in Science journal.
“But using auto research and a chunk of data, something was published internally that we said, oh my god. And that would normally be a PhD thesis that would take seven years.”
Friedberg says Auto Research accomplished in thirty minutes work that would normally take seven years as a PhD thesis.
“But the technical stuff, the science stuff, we did something in thirty minutes using auto research, and I love your view on auto research and what that tells us about how far we still have to go in terms of efficiency.”
Friedberg predicts AI apps will create ecosystems where advertisers pay to be embedded, offering dual revenue models.
“There could almost be, whether it's through kind of connectors or embedded tools, an incredible ecosystem that were traditionally advertisers actually pay to be embedded and show up inside of the AI app, and the consumer can either pay for it or the advertiser can pay for it.”
Friedberg argues no late-stage empire has ever gained a technology like AI before.
“There's never been an empire that's been in the late stage that suddenly got a technology like AI.”
Friedberg argues AI enabling humans to reach their potential is a fundamental right in education.
“Practically no one is reaching their potential. Yeah. What AI does is it allows humans to reach their true potential Yeah. As a human.”
Friedberg argues AI enables humans to reach their potential, a fundamental right in educational systems.
“What AI does is it allows humans to reach their true potential Yeah. As a human.”
Friedberg reports China now holds fusion plasma magnetic fields for thirty minutes using AI control systems.
“And they're now able to hold these magnetic fields for thirty minutes at a time out of China now”
Friedberg reports AI has increased fusion plasma stability from seventeen seconds to thirty minutes in three years.
“AI seems to be solving that problem because they're now using AI to train the control of the magnetic fields in a way that's working.”
Friedberg predicts wave of AI-native operators acquiring stale first-generation digital businesses to unlock AI opportunities.
“And when you take a look at those businesses as a modern day AI operator, you're like, what the hell? This thing is so underutilized.”
Sacks questions whether AI compute pricing at $30-50 per watt is sustainable, while Musk argues memory constraints will keep prices elevated.
“ But the segment that drove the stock's decline — the AI compute buildout — is also the one with the largest open questions, and Sacks sharpened them into two. First, is the $30-50 per watt spot price for compute durable? Elon argued the market is memory-constrained: memory production is growing perhaps 20% next year while demand is up 200% plus, so the bottleneck persists and prices may rise. Sacks's corroborating arithmetic: a gigawatt is a billion watts; at $50 per watt, a gigawatt is worth $50 billion in annualized revenue. ”
Friedberg characterizes data center infrastructure capex as high alpha and low beta, while model development is high alpha but very high beta and risky.
“I would frame it as capex is high alpha, low beta in data center infrastructure... model development theoretically could be high alpha but it's very high beta — it's a very risky way to deploy capital.”
Friedberg categorically rejects the idea that AI will destroy jobs, calling it a Luddite narrative disproven daily by what he observes on the ground.
“There is no job loss with AI. I’ll say it again. I’ve said it a thousand times, and I’ll say it again and again and again. The idea that AI is going to destroy jobs is a Luddite idea that is being disproven every single day. I see it on the ground. It is only a matter of time before people wake up to this and realize that this narrative that they’ve been sold is a crock of bleep.”
“And this has really been the heart of The US economy, at least for the last thirty years, probably going back even further.”
Friedberg says his company will continue using Salesforce products rather than building custom alternatives with AI coding tools.
“We're already on Slack. We're not going to go vibe-code Slack. We're not going to vibe-code CRM.”