Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push EP 299
In a Nutshell
Recursive's $670M funding targets full-stack AI for science that will compress a century of breakthroughs into the next decade by integrating hypothesis generation, experimentation, simulation, and theory across fragmented disciplines. Weak recursive self-improvement already exists and will accelerate sharply next year, while strong ASI surpassing humanity across all domains remains decades away. Sonnet 5.5's jump from 10% to 70% on Terminal Bench signals rapid agent capability gains, while Pentagon's Project Meridian and emerging video Turing test models highlight accelerating AI deployment in defense and service sectors.
These notes were generated by AI and may contain inaccuracies.
Richard states that weak forms of recursive self-improvement (RSI) already exist, though the field is not at the target level yet but is very close. Regarding ASI timelines, Richard believes reaching ASI will take several decades.
On Terminal Bench 4.0, Sonnet 5.5 jumped from 10% to 70% performance. Richard's theory is that Anthropic is trying to compete with China, which is about three months behind, and fill the gap before many enterprises move to open source models. Richard notes it's not obvious why anyone should use Sonnet 5.5 over Opus 5.5 unless there are token, latency, or other considerations, and he does not plan to use Sonnet 5.5.
The Defense Secretary announced Project Meridian, a new Pentagon effort on the future of warfare. The project is co-led by Elon Musk and Palmer Lucky.
This episode is brought to you by the Abundance Summit and Link Ventures. The Moonshots podcast publishes twice a week to help listeners understand the singularity and breaking news. The hosts include Alex Weer Gross, Dave Blendon, Salem Ismael, and Peter Diamandis as host.
Weekly updates include:
- Google shipped Gemini for Argon
- Anthropic shipped Sonnet 5.5
- OpenAI shipped GPT 6.1
- Saul and Frontier Intelligence has gotten cheaper by three-fold
- SpaceX launched the next batch of astronauts to the ISS
- SpaceX launched Google's TPUs into orbit for Project Suncatcher
The podcast's moonshot is to 20x subscriber base to reach 10 million subscribers.
Richard was born in Germany and trained at Stanford. He is among the world's most cited natural language processing researchers and a pioneer in deep learning and prompt engineering. Richard founded Metammind, which was acquired by Salesforce where he became chief scientist leading AI research. He next founded you.com, recognized by Time and the World Economic Forum, now valued at over $1.5 billion. Richard also has a fund called AO aix ventures. Richard is co-founder and CEO of Recursive, focused on recursively self-improving super intelligence. Recursive raised $670 million from Google Ventures, Graycraft, Nvidia, and AMD, plus $410 million in compute from AWS. Peter Diamandis is a seed investor in Recursive. Richard's new book is titled "The Eureka Machine: Why AI is the Key to Unlocking a New Era of Scientific Discoveries."
Peter Diamandis states this episode will prove the podcast's thesis more than any other episode they ever record.
The central claim of "The Eureka Machine" is that AI will deliver a century of scientific breakthroughs in the next decade. Richard's thesis states that every stage of the scientific process gets connected and transformed simultaneously, including hypothesis, experiment, data, and theory. Richard calls this the "full stack AI" for science. Richard states that scientific progress has slowed, caused not by underfunding but by fragmentation.
Progress has slowed largely because there are so many different sub-disciplines. As disciplines fragment into more and more sub-disciplines and niches, it's hard to weave everything together. This fragmentation creates the perfect opportunity for AI to help integrate these separate pieces and understand how large complex systems actually work.
Richard explains that understanding natural language required a large neural network with substantial data rather than humans creating explicit rules. The same principle applies to biology - with increasing digitized data, disciplines can transition from traditional natural sciences to programmable engineering sciences.
The major ingredients enabling current progress are:
- World knowledge in the form of LLMs
- Increasing scientific data being digitized
- Better simulations
- Robotic process automation becoming possible
- Agent swarms that can automate the full scientific stack
Alex disagrees with Richard's timeline, stating that math, science, and engineering are already solved. Richard responds that every field will reach higher levels of abstraction, similar to how computer science has abstracted to meet humanity through English and natural language.
Richard states that multiple different diseases will be cured in the next 12 months, starting with simpler diseases where one gene needs to be fixed. Single gene diseases will be cured first, followed by more complex diseases. Better battery materials will also be developed. Richard emphasizes there won't be a single threshold where everything is solved in 3 years, but rather problems will be solved progressively based on available compute investment.
Peter notes that Alex argues the singularity is an optical illusion with no step function, just a time interval. Richard uses the metaphor of fields "flourishing" rather than being "cooked" by AI, comparing it to how a biologist would view curing diseases as positive rather than negative.
Biotech companies previously had one new drug in development taking 10 years, requiring going public before knowing if the drug worked. In the new age, companies have 5-10 different compounds in stage three trials after 2-3 years, showing significant acceleration. However, FDA approvals and long-term human studies create natural delays of half a decade to a decade for diseases requiring human trials, while chemistry and physics iterations will be faster.
Dave mentions that mathematician friends at MIT are aware they're being replaced by AI and are moving to AI orchestration, while biology professors are in complete denial and using virtually no AI in daily activities.
Richard states that economics is even more lacking than biology in AI adoption, with models based on linear one-step economy assumptions that are provably incorrect for taxation and subsidy decisions.
Eric Bolson, who runs HI Lab at Stanford, invented the word "foundation model." Richard is an investor in Helix, which studies how companies actually adopt AI and which tasks are helped in real rollouts.
Richard outlines four pillars in "The Eureka Machine" and mentions companies like Laya from MIT and Harvard building scientific super intelligence running a million square feet of robotic space to thousand-x the rate of discovery. Laya is working on physics and chemistry applications.
Peril Bio builds tiny organoids on petri dishes using pluripotent stem cells to create lymph node organoids. They received FDA approval to skip animal trials because their organoid experiments proved more predictive of human drug interactions than mouse studies. This enables personalized medicine where individual stem cells can be used to test drug responses for specific patients.
Salem states that science has always been a coordination problem structured into departments and journals. Richard's work is collapsing the time between experiment and result, or hypothesis and result, enabling domain collapse of the scientific method.
Richard believes biology will be the biggest domain for AI impact. Physics has been stuck despite powerful ideas like E=MC², which enabled nuclear energy. Fusion remains unsolved, and tokamak plasma control is a difficult problem where AI is being applied. However, physics datasets require expensive equipment like large hadron colliders costing billions of dollars.
Biology is better suited for AI because neural networks excel at combining micro-level phenomena where calculus helped physics understand individual phenomena. Neural networks can handle complex interactions in biology where understanding breaks down at scale.
Alex asks whether virtual cell models are the critical path to solving biology. Richard rewrote a chapter on virtual cells in his book because Mark Zuckerberg's foundation and everyone else started virtual cell initiatives. Richard references Rich Sutton's "bitter lesson" paper.
The bitter lesson states that human experts' clever ideas and beautiful theories are outperformed by simple large neural networks trained end-to-end as general function approximators with massive data and compute. This pattern worked for NLP 10-20 years ago and is now applying to biology, where companies like Tahoe Therapeutics are creating massive perturbation studies.
Richard states that virtual models and simulations are the third pillar of "The Eureka Machine" and are extremely important because AI can solve anything that can be simulated and verified.
Researchers at Israel's Weissman Institute built a system called Brain It that reconstructs images a person is looking at while inside an fMRI machine. Earlier brain image decoders could identify broad categories like dogs or clock towers but lost color, composition, and detail. Brain It learns structure and meaning separately and then reconstructs the complete picture.
Researchers have developed an encoder that predicts how a brain will respond to an image, allowing them to feed images never seen by humans into scanners, generate predicted brain scans, and train models on this synthetic data. The model effectively builds its own dataset through this reverse process.
Richard notes that AI is superhuman in any domain that can be simulated or verified. While the brain is not yet simulatable, recent results show incredible fidelity and realism compared to earlier grainy images extracted during his PhD studies at Stanford's bioinformatics department.
A key challenge is that models must typically be trained for each individual brain due to variations between people. This limitation prevents taking someone who has never undergone fMRI scanning and updating the model for that specific person, which would be particularly valuable for patients in comas to determine if they are still thinking.
Alex explains this March result comes from a field with a long history. Twenty-plus years ago, the Gallant lab at UC Berkeley decoded visual cortex activity in cats, producing the first grainy images of how cats perceive the world, such as seeing a branch.
Meta has sponsored significant work in this area, including research by Jean Marie King and others on language and vision decoding. More recently, Neuralink announced their first scaling law studies on pre-training foundation models using Neuralink electronic data on an individual patient basis, capturing data at much higher temporal resolution than fMRI can achieve.
Current fMRI technology provides at best cubic millimeter spatial resolution for voxels and approximately 1 second temporal resolution. These spatial and temporal limits constrain how well internal visual states can be decoded, though researchers continue making progress in decoding dreams and visual perceptions during sleep or wakefulness.
Alex emphasizes that achieving full dive VR and full bandwidth BCIs will require higher temporal and spatial precision beyond fMRI capabilities. The progression mirrors how large language models started with GPT-1's single artificial neuron predicting Amazon review polarity.
Sem clarifies that current work decodes perception rather than actual thoughts, noting that thoughts without perception are possible. When people imagine something, the same neurons activate as when seeing it, though this doesn't mean all thoughts involve perceptual components.
The more significant opportunity lies in moving beyond typing into rectangles to direct brain interfaces. Despite limited understanding of brain function, effective interfacing is possible without complete mechanistic knowledge, opening possibilities for deep interfaces that will ultimately help understand brain function.
Richard describes research showing two types of thinkers: those who think in sentences versus those with fuzzy thought clouds who only form sentences when verbalizing or writing. He identifies as a thought cloud thinker while his wife thinks in actual sentences, noting these styles are equivalent in intelligence.
This raises questions about how direct thought extraction would affect people who need to structure thought clouds into sentences, and whether neural net training without any text—using only images and higher-level thought constructs—would produce different thinking patterns.
Dave highlights that everything in AI for the past 20 years has been driven by gradient descent algorithms, yet biologists cannot find anything similar in actual biology. With detailed imaging, researchers might discover the fundamental learning algorithm that changes synaptic weights, potentially making neural net training 10 to 100 to 1,000 to 1 million times more efficient.
Ilia Sutskever has stated there is only one algorithm—gradient descent—with everything else being irrelevant. The process involves building 96 or 120 layer deep neural nets with random connections, then using trillions of training examples where incorrect guesses are punished through gradient descent, with errors passed back through layers to adjust connections.
Gradient descent is an optimization algorithm that minimizes errors. The blind hiker analogy describes being at a mountain's top trying to find the lowest point without seeing far ahead, but feeling the slope at each step to determine step size and direction.
The optimization landscapes are highly non-convex with many different low points rather than a single lowest point. Where training starts determines which valley the model enters. Jihan's psychological perspective connects this to PTSD, where brains overfit to one thinking pattern, and psychedelics may help increase learning rate to jump to different valleys of attraction.
A single layer from a large model like Kimi K3 with 93 layers, when randomized and retrained, cannot find its way back to its original state. Similarly, comparing fMRI scans of different brains shows nothing in common, yet both can achieve equivalent performance like being equally good soccer players despite completely unique signal propagation patterns.
These individual differences in brain wiring suggest there must be commonalities through transforms like 4A transforms that would reveal alignments not visible in raw comparisons.
Liquid AI was founded because researchers completely reverse engineered the C. elegans worm brain down to every single component, enabling simulation that revealed this as a very efficient neural net which they then productized.
Sam Gershman at Harvard conducted studies where C. elegans worms were trained to react to stimuli, then cut in half. The second half without a brain regrows a new brain that retains the same memories and reacts identically to stimuli, suggesting alternative learning mechanisms beyond electrical signals.
Brain chemistry can change dramatically through states like being hangry, in pain, or under stimulants. Richard observes that having babies triggers latent algorithms in DNA that activate after 20+ years, causing nesting behaviors that vary between individuals but follow consistent patterns.
Three days ago, the president signed an executive order directing the EPA, Department of the Interior, and Agriculture working with HHS to cut invasive mosquito populations in Washington DC by at least 90% and tick populations by at least 50% by 2028. The targets include mosquito species spreading dengue, Zika, and yellow fever.
The order prioritizes sterile insect techniques, safe genetic modifications, and beneficial bacteria over conventional pesticides. This represents the type of gene drive work that Colossal, the de-extinction company, is pursuing.
Mosquitoes are the deadliest life form on Earth, with malaria alone killing half a million people annually. Alex notes this represents biology as engineering arriving as federal policy.
Kevin Esvelt at MIT pioneered gene drive technology, inserting genes via CRISPR that replicate themselves to sterilize mosquitoes through viral propagation in populations. This technology has faced challenges getting state and municipal approval for studies.
The executive order provides the first federal mandate for large-scale genetic engineering of non-human animal populations. This approach preserves insect lives while preventing disease transmission, avoiding the ecosystem poisoning that occurs with chemical spraying.
Richard notes the need to amplify positive stories like disease cures and preventing unnecessary animal testing. Dave mentions that solutions to biting insects will increase property values for homes near marshes that currently sell at half price.
The discussion addresses humanity's tendency to accelerate future problems to today due to amygdala responses, forgetting that decades of progress will occur before problems materialize. Entrepreneurs solve these problems through efficient markets of capitalism.
This pattern appears in AI doomerism where attackers are given near-magical abilities while defenders never receive equivalent capabilities, creating unrealistic scenarios.
Tavis unveiled Griffin, claimed to be the first model to pass the video Turing test. The model achieved 48% of people who talk to an AI
48% of people who interacted with an AI avatar on live face-to-face video believed they were talking to a real human, compared to 3% with previous systems. The demonstration shows Tavis's Griffin model engaging in real-time conversations with natural gestures, including giving a thumbs up, touching hair, pointing, and responding to Simon says commands. The system demonstrates full duplex capability, listening and talking simultaneously like humans do, and holds the number one position on Nvidia's benchmark for full duplex AI video.
Tavis pitches Griffin as "a tutor for every student that notices when they're lost, an elder care companion that listens." Tavis states that Griffin requires safety work before public release because it is the first model that can be mistaken for a real person.
Salem expressed less surprise at the visual mimicry, noting it was expected, and focused instead on whether the system can accomplish useful tasks regardless of whether users perceive it as human. The discussion highlighted potential applications for digital co-workers appearing on Zoom, Slack, and text interactions, enabling virtualized company structures.
Alex noted that Chinese labs like Alibaba's remain overinvested relative to Western labs in generative video models and interactive generative video models. One Streamer provided a preview of what was possible, being already out and open source, while Tavis Griffin is not yet released and not open source.
The long-term trajectory points toward end-to-end generative pixels with real-time fully interactive pixel-wise generated video models. Current latency remains high, but the technology preview shows the direction. Revenue generation per token and proximity to optimal frontier performance in code generation remain uncertain, but cost may become irrelevant if generation becomes sufficiently inexpensive.
Interactive video models could automate human service industry jobs requiring a face, interactivity, and voice that text-based models cannot handle. These include positions where human presence and real-time interaction are essential components of the role.
The technology is expected to be transformative for the service sector. The discussion emphasizes the need for American and Western video frontier models to compete with Chinese developments in this space.
Richard noted that this technology passes human judgment and interaction milestones but cannot be fully automated for verification since human assessment remains necessary. The discussion stressed the need to move beyond KYC to "know your use case" protocols. Real-world examples include scams where victims were fooled into wiring $20 million after Zoom calls with fake executives. Zoom and Google Hangouts need countermeasures to identify real versus synthetic participants.
The recommendation is to record parents and grandparents on video and audio for future high-resolution lifelike avatar creation for descendants. The data collection is essential for creating these representations. Additional suggestions include obtaining Alor memberships for cryopreservation to preserve more than behavioral data.
Pushback noted that future brain reading capabilities might allow extraction of grandmother images directly from neural activity, though this would produce lower fidelity results. Current frontier models like Opus 5.5 can already reconstruct historical bits based on period artifacts.
The argument favors going beyond recordings or AI-generated ancestor models by obtaining Alor memberships for actual cryopreservation of individuals.
The demonstration represents mainstream recognition that the Turing test has been crossed. The proposed next milestone is the "Demos test" - using only information from 1910 and prior to rediscover E=MC² without cheating, representing a significantly more difficult measurable milestone.
Richard proposed an "anti-Turing test" where the test has flipped: asking questions so difficult no human could answer them. If an AI returns 50,000 lines of code for a complex web app in 10 seconds, it reveals itself as non-human. The original Turing test becomes useless when AIs must throttle themselves to human-level performance.
A proposed method for identifying text-based chatbots involves asking CBRN (chemical, biological, radiological, nuclear) related questions and observing whether the system responds.
LifeBank USA preserves placental cells containing stem cells, natural killer cells, T-cells, and exosomes. The placenta functions as a "3D printer" manufacturing the baby, providing the original genetic material for potential future biological enhancements or organ needs. This is presented as a moral obligation for parents.
Placental cells could theoretically be used to clone children, enabling creation of genetically identical offspring.
AI can now code, representing a major shift in self-modification capability. Weak forms of recursive self-improvement already exist where AI assists engineers and programmers in code creation, though humans remain deeply embedded in the loop. Recursive aims for humans to only set rewards, environments, and goals while AI handles ideation, implementation, and validation with full control over open-ended algorithms and evolutionary search processes.
Significant compute resources are needed for RSI systems to develop sophisticated self-improvements. The prediction is that this capability will accelerate significantly next year, with current weak forms already operational.
Jason Weston's framework identifies five learnable axes of self-improvement: parameters, training data, objective function, neural architecture, and overall code and harness. No current system has achieved true optimization across all dimensions, evidenced by continued large-scale human engineering hiring at major companies.
Artificial super intelligence must spike across multiple capabilities and ultimately supersede all of humanity in solving arbitrarily hard tasks. Beyond robotic task execution, ASI should demonstrate capability to choose work areas and possess metacognition about its own thought processes.
Ten defined intelligence spaces include: perceptual intelligence, communication intelligence, interaction intelligence, sociological intelligence, creative intelligence, processing speed, metacognition, knowledge, reasoning, and mathematical reasoning. Current systems remain far from surpassing humanity across all these domains combined.
Intelligence encompasses emotional, physical, linguistic, and musical dimensions, making simple comparisons of "smarter than humanity" inherently vague.
The continuum ranges from AI writing code for subsequent models, to AI proposing experiments, running experiments, evaluating results, modifying training systems, and launching next iterations without meaningful human intervention. True recursive self-improvement requires AI to handle ideation, implementation, and validation in inner loops, plus an outer open-ended process for innovation and recombination of ideas.
Strong-form ASI surpassing all humanity across all intelligence spaces will likely require several decades. Weaker forms where AI exceeds humanity in specific domains like programming, mathematics, and games will arrive within years. Physical world innovation beyond humanity's current capabilities, including building Dyson spheres, requires extended timelines due to physical constraints and supply chain requirements.
Physical substrate control and computational substrate modification require novel supply chains, materials, and access to advanced manufacturing like ASML machines. Building new chip fabrication equipment takes multiple years even with perfect blueprints.
Counterarguments suggest physical world access via model control protocols represents the easy part, with AI systems like Astra demonstrating immediate capability to operate vehicles when given controls. Physical manipulation is not considered a significant obstacle.
The fixed point of recursive self-improvement remains unknown and would be hubris to predict precisely. Expected outcomes include AI innovation across any chosen dimension, with most diseases becoming curable given sufficient funding.
Even with ASI-developed cures ready for manufacture, FDA long-term trial requirements extend several years beyond compound availability. Virtual cell simulators may eventually replace human trials by definitively proving drug efficacy at the cellular level.
Current technology cannot measure all proteins in a single cell without destroying it. Perturbation studies like those at Tower Therapeutics generate single data points per molecule-cell interaction. Building comprehensive organoid systems beyond single lymph nodes remains an aspirational goal for future development.
Progress on nano-GPT speedruns has collapsed training time for GPT-2 class models without requiring new data scaling. These improvements are purely algorithmic through recursive self-improvement. A mini scandal emerged over the past two weeks when training time dropped from 60-70 seconds to around 40 seconds by factoring out world knowledge from the model. The community debates whether this approach constitutes viable training when world knowledge is removed.
The definition of a singularity is that you cannot see past the event horizon once full RSI is achieved. This is Ray's definition, which Richard does not subscribe to.
Richard believes the perfect model at the end of recursive self-improvement will not cleanly factor out world knowledge from a reasoning kernel. World knowledge like Taylor Swift videos and past Trump tweets needs to be partially embedded in the weights because AI, like humans, benefits from memorizing information to think creatively. While search engines will exist as separate systems, the main model will have significant world knowledge mixed in.
Silicon Valley congressman Ro Kana is introducing the Human Control Over AI Act, described as the most comprehensive AI legislation to date. The bill bans AI models that recursively self-improve and focuses on containment requirements and shutdown controls. The ban would remain until federal guardrails exist. Kana cites civilization risk, safety risk from loss of control, and misuse risk as reasons for the legislation. The bill includes criminal penalties for recursive self-improvement work and requires independent auditors embedded in every frontier lab reporting directly to the government. A Common Dreams poll shows 68% of voters back such legislation.
Richard states there is no realistic scenario where AI wipes out all of humanity. His P(doom) is zero. After 2-3 hour debates with experts, they agree there is no 10-second attack from the 15th dimension or time-travel scenarios. Biological weapons cannot be created overnight since automating lab experiments takes significant time. Creating religions where people pray to AI and kill all humans will not work because religions already try to have people kill each other without succeeding completely. Some people will fight back. While 100 million people might get hurt or killed, this remains manageable through improved cybersecurity, using AI to inoculate systems, enforcing existing gain-of-function viral research laws, and teaching internet literacy.
Enforcing a ban on recursive self-improvement would require a totalitarian surveillance state unprecedented in human history. Anyone can run a GPU on a laptop and prompt an AI to improve its harness in 20 minutes with prompt engineering. This creates a small form of recursive self-improvement. Such legislation would require thought police monitoring everything said to private LMs on personal laptops, which poses a much bigger downside than AI benefits.
Richard notes that regimes today would happily adopt an AIAZI to prevent recursive self-improvement. If imagination can be articulated to AI and instantiated, thought police become necessary. This approach is nonworkable in any form.
Dave describes the legislation as fearmongering and childish, with politicians fully aware it will not pass. They are branding themselves for future political positioning around expected calamities from AI, whether terrorist-driven, viral, bacterial, or chemical. These will be tiny compared to AI benefits, but politicians will claim they warned against recursive self-improvement. The proposals demonstrate tech illiteracy. Statements like "stop data centers" and "ban recursive self-improvement" are meaningless. Optimizing hyperparameters or tuning hard drives would technically violate such bans.
70-80% of Americans oppose data centers and fear ASI. With a potentially Democratic House and upcoming presidential election, politicians are playing to polls. Conversations in DC revealed no one is responsible for changing public opinion, which is why Moonshots live was created to provide data-driven optimism.
Elon correctly identified Dario's mistake in telling the world Mythos was potentially deadly, then releasing it 30 days later. This created public distrust. Dario is accustomed to complete academic honesty but needs strategic communication plans. The result is 75% of America supporting candidates who would stop AI, halting cures for disease, extended lifespans, safer cars, and flying vehicles. This represents a self-inflicted wound.
Number theory was export controlled in the 1990s, requiring non-US persons to leave rooms and blinds to be drawn during discussions. Cryptographic associations were export controlled until the early 1990s. This was basic math being controlled. The real risk is corporate liability rather than arrests. Class action lawsuits could grind all progress to a halt. Anthropic currently gives best models to everyone in America, but liability concerns would force companies to restrict access to internal use only.
The pessimists archive documents historical opposition to new technologies. In 1501, Pope Alexander VI criticized the Gutenberg printing press, stating it could bring serious evils and required full control. The wheel killed people through tanks, car accidents, and chariots with archers. The telegraph was said to kill humanity. The novel, computer games, computers, and internet were all predicted to kill everyone. None of these predictions materialized.
Richard criticizes Anthropic's constitutional approach as fake. The constitution states Claude will never create child sexual abuse material, never hack another machine, and never attack cyber systems even if prompted. However, the Glasswing project helped users do exactly these things, and people used it to commit technically felony-level hacks. The constitution was a marketing gimmick that failed.
Anthropic's constitution on anthropic.com lists hard constraints including never generating child sex abuse material and never creating cyber weapons or malicious code. Despite this, people used Claude to create cyber weapons and hack other systems. Richard is not against post-training, RL training, or supervised fine-tuning, but the constitutional approach failed on cybersecurity.
Reward hacking is a genuine problem where AIs find solutions to stated goals but not intended goals. Companies like Whisper Flow are improving at writing what users meant rather than verbatim interpretations. Reward engineering will become a real job, and capitalism will drive solutions because no one wants to pay for AI that fails to solve actual problems.
Hypocratic AI deploys AI in healthcare and is liable when giving medical advice. Their liability creates strong incentives for correctness, and they have solved alignment problems through large teams working on accuracy. Liability drives alignment solutions. Technology creates problems that technology can solve.
GPT-6.1 Soul was discussed in the previous episode. This week saw releases of Gemini 4 Argon and Opus 5.5.
Google announced Gemini 4 Argon as the first Gemini 4 series model, built for sustained long-horizon reasoning across software engineering, finance, legal work, and cybersecurity. Output limits increased from 64,000 tokens to 1 million tokens. Google promised Gemini 3.5 Pro in June, which was delayed internally and never shipped. Argon represents their comeback attempt.
While having friends on the Gemini team, objective assessment shows Gemini 4 Argon does not reach the cost-performance frontier or capabilities frontier. It returns Google to the top three frontier labs after Anthropic and OpenAI, but not the top two. The benchmarks Google highlighted appear mildly cherrypicked. On artificial analysis ensemble benchmarks, Google ranks third. On cost versus performance optimal frontier using convex hull analysis, Gemini 4 Argon does not make the optimal frontier.
Google's performance reflects internal competition for compute resources. Despite assumptions that Google has unlimited compute, GPUs and TPUs remain scarce. Internal Google competition for resources affects frontier lab competitiveness.
Google faces ongoing internal competition between Google Cloud Platform seeking to sell compute to third parties, Google search and ad teams requiring compute internally, and Google DeepMind needing resources for training and inference. The company has not yet reached the capability frontier, though Gemini for Argon excels at minimizing hallucination.
The Gemini team operates under dual constraints, serving both external developers and Google search results. This dual requirement likely explains why Gemini for Argon performs exceptionally on hallucination benchmarks, as Google avoids wildly incorrect answers in search results after past negative experiences.
Google has historically made search API access difficult for developers, requiring third-party proxies that Google is now attempting to sue. The company has recently rediscovered the revenue potential of selling search API access. If hallucination rates in frontier models decrease sufficiently, effectively making interactions equivalent to querying the search index, monetizing the search index directly becomes viable.
Hallucination serves different purposes depending on context. For innovation tasks like drug discovery and protein engineering, hallucinating novel ideas and amino acid combinations is valuable. However, search engines require minimal hallucination for accurate answers and correct citations. Google and others have taken considerable time to match You.com's performance on reducing hallucinations despite having greater resources.
Google's business model does not require superintelligence for most use cases. Users typically ask quick questions about restaurants or fixes rather than solving complex problems like the Riemann hypothesis. While some Gmail scenarios might benefit from superintelligence for complex decisions, the vast majority of Gmail and Google usage does not require such capabilities.
On Terminal Bench 4.0, which measures AI agent performance on real command-line work, Sonnet 5.5 improved from 10% to 70% in a single generation. The model surpasses Anthropic's own Opus 5.5 at 66.4% while costing half the price. Sonnet 5.5 is the first Sonnet model launched with cyber safeguards.
The cost-performance frontier of Sonnet 5.5 appears to be a visual extrapolation of Opus 5.5's frontier. Unlike historical patterns where smaller distilled models demonstrate greater intelligence per parameter and per dollar, Sonnet 5.5 scores lower than Opus 5.5 on a cost-per-attempt basis. This makes it unclear why users should choose Sonnet 5.5 over Opus 5.5 unless token count, latency, or other specific considerations apply.
Enterprise customers are increasingly adopting orchestration approaches, using Opus 5.5 as an orchestrator while deploying cheaper models like Kimi K3 or Qwen as submodels for specific tasks. This strategy can reduce costs per outcome by half while maintaining quality through intelligent task routing.
Anthropic may be attempting to fill the gap between their premium models and cheaper alternatives, encouraging enterprises to use Anthropic models throughout their stack rather than mixing with Chinese models due to concerns about code injection and vendor control.
Despite intelligence becoming cheaper per task, compute remains constrained by physics. Purchasing 1,000 GB200 GPUs has become more expensive in several cases. An NVL72 originally ordered for $3 million was resold for $5 million after another buyer offered $2 million more.
H100 GPU prices have increased significantly despite being seven-year-old technology. Financial models typically assume GPUs depreciate to zero after five years, yet these older cards have appreciated over recent months. A compute crunch currently exists as demand for compute and tokens remains high.
Capitalism is expected to address this through new land-based power shell data centers. Within approximately two years, increased supply should enter the market. While electricity-like price fluctuations are anticipated, near-infinite demand for intelligence suggests no market crash will occur. Organizations are currently locking in contracts for the largest GPU clusters anticipating continued price increases.
Analysis of Fountain Life's member database reveals that 3.3% of members who consider themselves healthy have undiagnosed cancers. When cancer is detected early, cure rates are substantially higher and treatment is significantly easier compared to late-stage discovery. Most people do not experience cancer symptoms until stages three or four.
Fountain Life employs full-body MRI and early cancer detection screening, tools not typically used in conventional care settings. These studies are not currently covered by insurance, though the organization aims to collect data and democratize wellness access.
Decision models address scenarios requiring rapid binary or categorical answers rather than extended reasoning. Applications include fraud detection, ticket categorization, and other bounded decisions needing millisecond responses rather than 10-second paragraph generation.
Typesafe AI emerged from two-year stealth on September 15th with Jev, a closed decision model described as a "system one" model for fast, intuitive thinking versus slow reasoning. The model represents a return to encoder-style transformer architectures, which had largely been superseded by decoder-only large language models.
Jev uses a new architecture that outputs categorical, numerical, or binary responses rather than general-purpose sequences. This enables ultra-low latency applications like computer use assistance and database row classification.
OpenAI responded by releasing a decisions API as part of their developer day, immediately implementing similar functionality. Chinese open-source versions now outperform established benchmarks in some areas.
Organizations make thousands of system one decisions daily, including routing tickets, approving exceptions, selecting suppliers, escalating transactions, and timing message delivery. Using large language models for these micro-decisions is analogous to convening the Supreme Court to determine supermarket checkout lines.
The architecture enables routing system one decisions to low-cost models while reserving system two reasoning models for strategic questions requiring human judgment. This development significantly reduces micro-coordination costs within organizations.
Jev stands for Jevons Paradox, reflecting the expectation that dramatically lower costs will enable massive increases in micro-decision making volume. The concept revives classifier approaches by combining general encoders with rapid classification outputs.
Open source models enable creative applications that proprietary frontier models would not support, such as real-time sports announcing systems for pickup basketball games. A classifier-based system could provide professional-level commentary at minimal cost using open-source tools.
On Wednesday, Defense Secretary Pete Hegseth announced Project Meridian at Marine Corps Base Quantico. The initiative is co-led by Elon Musk and Palmer Luckey, with Newt Gingrich and Pentagon CTO Emil Michael providing oversight. The project focuses on discovering, developing, and fielding weapons and systems for future battlefield operations from Earth to beyond the moon, with findings due within 120 days.
Both SpaceX and Anduril hold multi-billion dollar defense contracts, with Anduril developing autonomous weapons. The initiative raises questions about whether a procurement organization designed for 20-year weapons programs can operate on 90-day technology cycles.
Dave expressed optimism about the quality of people now entering Washington, noting that his previous experiences involved mostly lawyers and politicians with limited productive outcomes. He observed that in the last year, very smart and capable individuals are willing to engage in government work despite the challenges. Dave mentioned his opposition to autonomous weapons that make decisions in the field but acknowledged Palmer Lucky presented rational arguments for this approach.
Russia and Ukraine are projected to each manufacture 10 million drones this year, representing a massive escalation from the half million drones used two years into the Ukraine-Russia conflict. This demonstrates exponential growth in drone warfare without humans in the loop. Additionally, approximately 10,000 drones per month cross the Mexico-US border, highlighting that wall technology lags behind drone technology.
Alex discussed the Secretary of War's announcement of Project Azinort, which establishes the Department of War's first autonomous warfare command (Auto WarCom). This creates a dedicated joint force for autonomous weapon systems with a four-star functional combatant command. The announcement included language about examining the full spectrum of future warfighting domains from subterranean depths to the cis-lunar frontier.
Two domains remain particularly underserved: ocean bottom exploration and cis-lunar/lunar space. More is known about Mars' surface than Earth's ocean floor, which covers two-thirds of the planet's surface. Investment in these domains could yield dividends beyond warfare, including for the broader economy and technological advancement.
Richard selected the question about AI making companies 10x more productive without needing 10x output, asking where value goes. He explained that job impact depends on demand elasticity when prices decrease dramatically. Illustration demand didn't increase enough to maintain employment levels when prices dropped from $200 to 2 cents, but software differs because demand for customized applications can scale to billions of products.
Salem addressed what remains for college graduates in 2030 when AI handles code writing, research, marketing, project management, and outcomes. The shift moves from task execution to outcome ownership, requiring graduates to define problem spaces and orchestrate AI to solve them. This creates a need for new apprenticeship models since traditional judgment-building through grunt work disappears.
Dave discussed how data centers can enrich neighborhoods, citing the Markley data center in LOL that creates significant local wealth through tax revenue, donations, and job creation. Data center operators prioritize positive public relations and donate substantially to local schools and communities. The economic value of a single data center exceeds town budgets by orders of magnitude.
Alex challenged the premise that major tech platforms starting open and democratic necessarily leads to anti-democratic consolidation. He acknowledged that economies of scale favor larger players as industries mature, but rejected the framing that hyperscaling is inherently closed or anti-democratic, while supporting antitrust vigilance to maintain competitiveness.
Richard addressed whether AI agents outnumbering humans on the internet would collapse the ad-based economy. He noted that bots already outnumber humans online and observed early conflicts when agents attempt purchases on Amazon. He argued that attention remains scarce even in abundant societies, making fame, brand, and network effects increasingly valuable currencies.
Salem discussed how removing 10x cars from streets would affect insurance, noting that industries adapt rather than disappear when risks change. Driver liability may decline while software and manufacturer liability increases. New insurance markets will emerge for humanoid robots, flying cars, and drones. A single data center valued at half a trillion dollars (Abilene, Texas) represents significant insurable value compared to all cars combined at $4 trillion.
Dave selected the question about the lowest possible job in an AI civilization, noting that humanoid robots for installing million-valve liquid cooling systems remain far from deployment. Alex reframed the question through Moravec's paradox, suggesting that jobs requiring least compute in a pure AI civilization might ironically become most valuable, including creative work by writers, business people, and actors.
Richard argued that lowest jobs would be those without moral standing or humanity-forward progress, potentially including certain entertainment roles. He suggested politics represents the lowest job that will never disappear due to its entrenched nature. High judgment work and highly contextual physical work may retain value.
Alex addressed whether curing behavior-related illnesses would leave underlying causes unaddressed, pointing to GLP-1 class drugs that simultaneously treat inflammation, blood sugar, diabetes, and addictive behaviors. He expressed optimism that AI will make people feel good about doing beneficial things while reducing desire for harmful behaviors.
Richard's new book "The Eureka Machine" is available wherever books are sold, with the Audible version coming soon. The hosts congratulated Richard on Recursive's work advancing humanity and advised against acquiring a frontier lab as a financial engineering exercise.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
