AI Costs Plummeting 40x: Why Costs Are Collapsing & What It Really Means w/ Dave, Salim & AWG

Anthropic Overtakes OpenAI in Enterprise LLM API Market Share [00:39]

• A shift in the Large Language Model (LLM) API market share is observed, with OpenAI's share declining and Anthropic's rising.
• The critical question is whether code generation is the key to recursive self-improvement for AI.
• If code generation is the path, Anthropic's strategy may lead to significant advancements, but if it lacks a crucial component for superintelligence, this trend might not last.
• Another perspective suggests that AI's business model is becoming clear, with enterprises trusting Anthropic for sensitive data while OpenAI focuses on consumers.

"I I I think the central question, Peter, is whether code generation is the critical path to recursive self-improvement."

Anthropic Projects $70 Billion in Revenue by 2028 [07:00]

• Anthropic forecasts $70 billion in revenue and $17 billion in cash flow by 2028, indicating significant growth in the enterprise AI market.
• While OpenAI and XAI have dominated headlines, Anthropic is gaining ground by focusing on the enterprise sector, which is seen as a wide-open market.
• Enterprises are increasingly using AI as a management tool, and Anthropic's Claude is a popular choice for backbone services.
• Google views Anthropic as a "friendly AI company" compared to the competitive landscape between OpenAI, Google, and XAI.

"The enterprise market uh is is wide open. And uh you know, what OpenAI is doing is directly competing with Google, which is very aggressive, very cool, but also risky."

Worldcoin's Marble Model Unveiled [12:40]

• Worldcoin's World Labs has unveiled AI models capable of generating entire virtual worlds, showcasing an approach using 3D Gaussian splats (3D GSs).
• 3D GSs are efficient, allowing for dynamic visualization on local computers, unlike pixel-wise generation which requires serverside GPUs.
• This development signals an emerging trade-off between compute efficiency and versatility in generating immersive virtual environments.
• The potential applications extend beyond consumer use cases to generating synthetic data for training vision-language-action models for robots and scientific discovery.

"It's generating not individual pixels but so-called 3D Gaussian splats or 3D GS's which are sort of transparent blobs that cumulatively together build up into what looks like a photorealistic 3D traversible world."

AI Can Now Forget Memorized Data Without Losing Reasoning [17:59]

• A new method allows AI to forget specific memorized data without retraining from scratch or compromising its intelligence and reasoning skills.
• This technique helps distinguish weights representing knowledge from those representing general reasoning capabilities by analyzing their impact on model loss.
• The primary benefit is improved generalization and efficiency, rather than solely enterprise data privacy.
• The ultimate goal is to achieve highly intelligent, smaller models, potentially with sub-billion parameters.

"But that really the emphasis in the inflection is less about some sort of enterprise privacy or some sort of like ML data firewall feature and more about figuring out which parts of the model and models are huge."

Google Introduces Nested Learning for Continual Learning [23:40]

• Google's nested learning paradigm focuses on higher-order meta-learning, aiming for "learning to learn to learn."
• This approach suggests that model architectures and optimizers might be two facets of a common paradigm, potentially leading to a grand unified theory of machine learning.
• The core insight is that machine learning can be viewed as information compression, where advanced intelligence emerges from sufficient compression.
• Nested learning enables AI systems to learn continuously, mimicking human-like lifelong learning.

"So, this is a a pretty dense, if I may say so, Nurup's paper. Uh, Nurup's the arguably premier AI machine learning conference."

Moonshot AI Launches Ultra-Low-Cost AI Model [28:08]

• Moonshot AI, backed by Alibaba, is launching a new ultra-low-cost AI model, significantly reducing the expense of training trillion-parameter models.
• The cost to train these models is estimated at $4.6-$6 million, making them accessible to many companies.
• This represents a 30-40x cost reduction compared to the original models developed by OpenAI and Anthropic.
• The availability of these advanced, open-source models from China is a major development for those wanting to experiment with AI's core.

"So, you know, these Kimmy models actually are top of Sweetbench, right up there with with Anthropic, you know, just 1% or so below."

Brussels to Loosen GDPR Rules for AI Boom [35:57]

• Brussels plans to relax General Data Protection Regulation (GDPR) rules to facilitate and boost the AI industry in Europe.
• CEOs and companies in Europe have expressed significant concern about their ability to compete due to data access limitations imposed by strict privacy regulations.
• The compliance burden for GDPR has been substantial, with AI audits costing around €260,000 and taking 8-15 months, delaying projects.
• This move aims to address the competitive disadvantage Europe faces against the US and other regions in AI development.

"You know, this is part of the problem with uh overregulation is you just slow this down. The cost of compliance for stuff like GDPR has proven to be a ridiculous thing."

Global Concerns: Cost of Living and Unemployment [41:36]

• A global survey reveals cost of living as the number one concern, followed closely by unemployment.
• Poverty and social inequities are the third major concern worldwide.
• In Africa, unemployment is the primary concern, while the rest of the world is more focused on the cost of living.
• This highlights the reality people are feeling, contrasting with discussions of future abundance and demonetization.

"So the number one concern globally is cost of living by far. Can we afford to live in this world? and and tied very closely to that is unemployment."

AI's Role in Solving Grand Challenges [01:21:23]

• Experts predict that AI will solve many grand challenges in math, science, engineering, and medicine within the next three years.
• The rapid advancement of AI models like GPT-5 and Gemini 3 is expected to significantly accelerate scientific discovery and patent filings.
• AI-generated patents and agent-to-agent intellectual property exchanges are emerging trends to monitor.
• The Chan Zuckerberg Initiative is investing heavily in AI for medical breakthroughs, aiming to cure diseases within the next five years, shifting focus from specific cures to generating more compute power.

"Science is going to get solved and all of its disciplines are going to get solved and AI is going to do it."

Genetically Engineered Babies: Banned but Pursued [33:30]

• Companies are emerging that are developing CRISPR capabilities for embryo editing, despite U.S. FDA bans on germline editing research and commercialization.
• Preventive and Manhattan Genomics are examples of companies looking outside the U.S., potentially to the UAE, to conduct this research.
• This marks a shift from genetic selection in IVF to genetic alteration.
• Concerns about eugenics are prominent, but proponents argue for starting children with the best possible genetic foundation, likening it to providing the best healthcare and education.

"So Wall Street Journal reported one out of San Francisco called Preventive that's backed with SAM. I mean Sam is backing an incredible number of companies, right?"

Blue Origin Lands Rocket Booster for First Time [01:13:47]

• Blue Origin successfully landed its New Glenn rocket booster on a recovery ship named "Jacqueline," named after Jeff Bezos's mother.
• This achievement is seen as crucial for increasing humanity's access to space and reducing dependence on SpaceX.
• It signifies a significant step towards reusable launch systems and provides a second major capability provider for space access.
• The success of Blue Origin's booster landing is a positive development for the growing space industry.

"I just think it's great that we have a second cap capability aside from SpaceX. I think it's good for the world."

Drones: Swarm Technology and Warfare [01:01:34]

• China has demonstrated a world record of 16,000 AI-powered drones flying in coordinated swarms, controlled by a smart AI system.
• This technology has implications for creating giant aerial screens and enabling complex construction or cleaning tasks.
• The warfare in Ukraine is heavily reliant on drones, with estimates of half a million drones deployed on both sides.
• Ukraine is poised to become a major drone manufacturing hub in Europe post-war.

"But I I really feel like the drone part of it or the the interactive thousands of drone part of it is way underappreciated as as a real world thing to do right now."

Elon Musk's Tesla and Flying Cars [01:04:54]

• Rumors suggest Tesla might unveil a flying car, potentially incorporating SpaceX's jet or rocket propulsion for hovering or hopping capabilities.
• This is seen as a "science project" that, like other ambitious ventures, replaces traditional marketing budgets and inspires innovation.
• While not a mass-produced flying car, this could revolutionize short-distance travel and accident avoidance.
• The development highlights Elon Musk's strategy of pursuing bold projects to drive progress and market cap.

"So, we've been talking about this is the next generation roadster. God knows what it's going to be charged and and this is not a uh you know a Model Y mass production."

Geoengineering and Climate Control [01:10:01]

• Elon Musk proposed a large, solar-powered AI satellite constellation to prevent global warming by making minor adjustments to solar energy reaching Earth.
• This concept, akin to a "solar sun shade," faces challenges related to the tragedy of the commons and requires global consensus.
• The idea of a global weather grid, capable of steering hurricanes and mitigating storms, is technically feasible with AI but politically complex.
• Geoengineering is seen as an inevitable response to climate change, as passive geoengineering through carbon emissions has already been occurring.

"So, Elon, how to prevent global warming. Quote, "A large solarpowered AI satellite constellation that would be able to prevent global warming by making tiny adjustments in how much solar energy reaches the Earth."

AI for Longevity and Healthspan [01:26:24]

• The US government is reducing the price of GLP-1 drugs, making them more accessible and demonstrating potential for universal basic healthcare services.
• GLP-1 drugs are showing remarkable efficacy across various biological dysfunctions, not just metabolic issues, and can halve the incidence of strokes.
• While not a panacea, these drugs are a significant step towards abundant healthspan, but require lifestyle changes like diet and exercise.
• The potential for AI to accelerate discoveries in longevity and disease cures, as seen with CZI's initiatives and the development of AI scientists like Edison's Cosmos, is immense.

"I I think offering for $150 odd dollars per month to to US persons who need it GLP-1 class drugs that starts to look like universally basically abundant health span drugs."