Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push EP 299

2h 25m 123K views 3 Oct 2026

Is it worth your time?

Worth your time if you want a detailed breakdown of how artificial intelligence is accelerating biological discoveries, advancing recursive self-improvement, and shaping defense and computing infrastructure.

This conversation explores how artificial intelligence is accelerating scientific discovery, reshaping biological engineering, and driving new commercial and defense applications. The speakers examine the technical and physical realities shaping the race toward advanced intelligence.

AI-driven scientific discovery. Socher argues that artificial intelligence will deliver a century of scientific breakthroughs within the next decade by connecting every stage of the research process. He points out that biotech companies are building tiny organoids on petri dishes, which secured Food and Drug Administration approval to skip animal trials because testing lymph nodes proves more predictive of real human body reactions than mice trials.

Recursive self-improvement and superintelligence. Socher explains that weak forms of recursive self-improvement are already operational because software can now assist with coding, though fully automated loops remain in development. He estimates it will take several decades to reach artificial superintelligence that surpasses human capability across ten distinct reasoning spaces, cautioning that giving an AI control over its own hardware requires entirely novel semiconductor supply chains.

Project Meridian and autonomous warfare. Socher warns that artificial intelligence should never control lethal decisions or nuclear codes without human oversight because autonomous warfare poses serious operational risks. The discussion details Project Meridian, a Pentagon initiative co-led by Musk to field systems for future troops across domains from underground environments to the cislunar frontier.

The compute crunch and smaller models. Socher points out that acquiring high-end hardware like GB200 clusters remains difficult due to fierce competition, causing prices for older GPUs like H100s to rise over recent months. He explains that organizations handle rapid routine tasks by routing them to low-cost classification systems like the Jev model, preserving expensive reasoning models for complex strategic questions.

Recursive Self-Improvement, ASI & What’s Ahead

Socher explains that weak forms of recursive self-improvement are already here, though true recursive self-improvement is not quite here yet while remaining very close. He estimates it will take several decades to reach artificial superintelligence. Discussing recent model performance, he notes that Sonnet 5.5 jumped from 10 percent to 70 percent on Terminal Bench, which measures agents on complex command-line tasks. He suggests Anthropic is trying to compete with China, which lags by about three months, and hopes to fill that gap before enterprises shift to open-source models. He sees little reason to use Sonnet 5.5 over Opus 5.5 unless facing specific token or latency limits, and states he does not plan to use it.

The Eureka Machine & AI-Driven Scientific Discovery

Socher begins by explaining the core premise of his book, The Eureka Machine, which posits that artificial intelligence will deliver a century of scientific breakthroughs within the next decade. He outlines the concept of full-stack AI, explaining that every stage of the scientific process—from hypothesis and experimentation to data and theory—gets connected and transformed simultaneously. He points out that scientific progress has slowed not because of underfunding, but due to fragmentation, as researchers are divided into narrow subdisciplines and niches, separating cell biology, molecular biology, and medicine. Artificial intelligence arrives to weave these separate pieces back together and help humanity understand complex systems, moving natural sciences from traditional observation into programmable engineering sciences.

He outlines the major ingredients driving this transformation: world knowledge in the form of language models, digitized scientific data, better simulations, robotic process automation, and agent swarms that will automate the entire scientific stack. Comparing this to computer science, he notes how abstraction has progressed so that people no longer program in zeros and ones or C++, meeting humanity at the level of English and natural language. When discussing timelines, he avoids a single threshold where everything is solved at once, predicting instead that simpler single-gene diseases will be cured within the next twelve months, followed by more complex diseases and better battery materials, depending on how much money is invested in compute.

Addressing the state of mathematics, he rejects the idea that the field is cooked, arguing instead that it is flourishing, though he notes that pure intellectual exercises without real-life impact are being left behind. For timelines in medicine, he explains that while biotech companies are accelerating trials with five to ten compounds in stage three trials after two or three years, human trials in the United States face natural delays and regulatory approvals from the Food and Drug Administration, making a half-decade to a decade the realistic timeframe for complex diseases, whereas chemistry and physics can iterate faster without long-term human trials. He agrees with the host's phrasing of AI flourishing mathematics rather than cooking it, and notes that mathematicians are shifting toward AI orchestration.

Socher turns to economics, arguing that it is even more lacking than biology, continuing to use linear models and lagging behind in adopting artificial intelligence for policy decisions. He praises Eric Brynjolfsson and his work with Work Helix on company adoption of AI, and mentions his investment in Laya Science, which is building a scientific super intelligence operating a million square foot of robotic space to accelerate discovery. He describes the scientific method as ideation, implementation, and validation, and notes the importance of open-ended swarms that can evolve and combine ideas, paired with wetware experiments.

Highlighting another investment, Peril Bio, he describes how they build tiny organoids on petri dishes using pluripotent stem cells. They secured Food and Drug Administration approval to skip animal trials because testing organoids, such as lymph nodes for immunotherapy, proves more predictive of real human body reactions than mice trials. He notes that physics is stuck on expensive data collection methods requiring large hadron colliders, making biology a better fit for artificial intelligence because neural networks excel at combining complex, interacting micro-phenomena like bacteria and neurons, just as calculus helped physics.

Addressing virtual cells, Socher connects them to Richard Sutton's bitter lesson, which demonstrates that simple methods utilizing large neural networks trained end-to-end with massive data and compute outperform clever expert hacks. Biology currently mirrors the past state of natural language processing, where experts doubted a single model could handle conversation, but companies like Tahoe Therapeutics are now conducting massive perturbation studies to gather data. He concludes that virtual models and simulations form the third pillar of his book and will play a crucial part in solving biology and disease.

AI Decodes the Brain & the Future of BCIs

The discussion turns to brain decoding and neural interfaces, starting with a study from the Weizmann Institute using a system called Brain-It to reconstruct images a person is looking at inside an fMRI machine by learning structure and meaning separately. The system can also run in reverse to generate predicted brain scans from unseen images, though Socher notes that models currently require individual training for each unique brain. The conversation notes a long history of visual decoding from the UC Berkeley lab to Meta and Neuralink, which recently announced scaling law studies on foundation models using electronic data at high temporal resolution. While fMRI has spatial limits of about a cubic millimeter and temporal limits of about one second, researchers continue to extract dreams and hallucinations, though full-bandwidth brain-computer interfaces will require higher precision. Socher distinguishes between decoding perception and reading actual thoughts, while the host points out that imagining something lights up the same neurons as seeing it, offering a potential interface to replace typing and help science understand how the brain works. Socher adds that people differ in whether they think in sentences or fuzzy thought clouds, and wonders what neural net training would look like without language. The group discusses learning algorithms, noting that AI relies on gradient descent while biology has not found anything similar, though detailed imaging may eventually reveal the fundamental learning algorithm. Gradient descent is defined as an optimization algorithm that minimizes errors, compared to a blind hiker feeling the slope to find the lowest point in a non-convex landscape, which Jihan compares to psychological states like PTSD where the brain gets stuck in a valley of thought.

The conversation moves to brain simulations, noting that Liquid AI reverse engineered the C. elegans worm brain to build an efficient neural net product. Socher mentions Harvard neuroscience professor Sam Gershman, who studied C. elegans worms by training them, splitting them in half, and observing that the half without a brain regrew a new brain that retained the same memories and reactions. He suggests that biology relies on more than electrical signals alone, pointing out how brain chemistry shifts with hunger, pain, or stimulants, and how DNA triggers latent algorithms like nesting after a baby is born.

Gene Editing, Mosquitoes & Engineering Biology

The host brings up a recent executive order signed by the president, which directs federal agencies to reduce invasive mosquito populations in Washington DC by ninety percent and ticks by fifty percent by the year 2028, targeting species that spread diseases like dengue, zika, and yellow fever. He highlights that the order prioritizes sterile insect techniques, safe genetic modifications, and beneficial bacteria over traditional chemical pesticides, noting that mosquitoes are the deadliest life form on Earth and that malaria kills half a million people annually. The guest adds that this represents biological engineering arriving as federal policy, recalling how an MIT classmate named Kevin Esfelt pioneered gene drive technology using CRISPR to sterilize mosquitoes, but faced immense difficulty getting state and municipal approval. He notes that the technology previously lacked a federal mandate for large-scale genetic engineering of animal populations, which this executive action now provides. He emphasizes that this approach avoids killing the insects directly, allowing them to live while blocking disease transmission, which protects ecosystems better than mass chemical spraying.

The guest responds that these positive stories need much broader public amplification because the future requires better marketing, comparing it to saving animal lives through organoid testing at Peril Bio to avoid animal dissection. The host remarks on the potential real estate impact, suggesting that eliminating biting insects will dramatically raise the value of marshland properties. Another speaker points out the importance of keeping bees alive and ensuring birds have enough insects to feed on, leading the host to argue that humanity has spent decades scared of its own shadow, missing out on decades of nuclear power progress due to misplaced fears from films like The China Syndrome. He draws a parallel to geoengineering, expressing hope that society is finally moving past excessive worry over its own shadow. The guest agrees, noting that well-intentioned alarmism can unfortunately push humanity away from major achievements like nuclear energy abundance, and refutes overpopulation myths by pointing to resources like humanprogress.org, which show that resources become cheaper as populations grow.

The conversation touches on Germany, where the guest notes an unfortunate trend of wanting to offramp from progress due to a degrowth mindset within environmentalism that views energy consumption as inherently harmful. The host observes that humans possess a strong tendency to spot a distant future problem, amplify it immediately through the lens of the amygdala, and overlook the decade of technological progress that will occur before that problem ever arrives. He stresses that entrepreneurs solve these challenges continuously, creating a forward-propagating solution set driven by capitalism. The guest connects this pattern to AI doomerism, pointing out how critics imagine attackers gaining near-magical capabilities while assuming defenders never acquire equivalent defensive advantages, constructing unrealistic scenarios about undetectable super-viruses with remote kill switches while ignoring the corresponding potential for equally advanced, rapid countermeasures like universal vaccines.

Tavus, the Turing Test & AI Avatars

Socher examines a company called Tavus unveiling Griffin, which it calls the first model to pass the video Turing test, noting that 48 percent of people speaking with an AI avatar on a live face-to-face video believed they were talking to a human, up from three percent in previous systems. The guest observes that while the mimicry is expected, the real question is whether it can perform useful tasks so that users stop caring whether the entity is human or not, pointing toward digital coworkers on Zoom, Slack, and text. He adds that Chinese labs remain overinvested in generative and interactive video models compared to Western labs, citing Wan-2.1 as a preview, and predicts that the long-term destination is fully generated real-time interactive video models where users can create scenes with people. He questions how revenue-generating per token such models are for code generation, but suggests that if generating arbitrary humans for meetings becomes absurdly inexpensive, cost-performance optimal frontiers matter less. He points out that many human service industry jobs requiring a face, voice, and interactivity could be automated away, though he warns that security countermeasures will be needed on platforms like Zoom and Google Hangout to prevent sophisticated fraud where AI impersonates executives to authorize wire transfers. Socher rejects the idea that passing this video Turing test constitutes a complete scientific milestone for automated verification, explaining that scaling remains difficult when a human must stay in the loop to confirm identity. Responding to suggestions about preserving family members as avatars, Socher and the other speakers discuss recording deep video and audio histories of parents and grandparents, while joking about cryopreservation and saving placental cells to clone children or build ancestor simulations. Shifting to benchmarks, Socher proposes that the Turing test has inverted, arguing that artificial intelligence now must throttle itself down to human levels to pass, whereas proving non-human intelligence actually requires asking questions too difficult for any human to answer, such as generating massive amounts of complex code instantly or tackling complex chemical, biological, radiological, or nuclear queries.

Recursive Self-Improvement & the Road to ASI

Socher explains that weak forms of recursive self-improvement are already here because AI can now code, allowing humans to close the loop on changing the software, though engineers still remain heavily involved. At his company, Recursive, the goal is to limit human involvement to setting up rewards, environments, and goals, leaving the entire process of ideation, implementation, and validation to the artificial intelligence alongside open-ended evolutionary search algorithms. While physical reality demands massive amounts of compute to train sophisticated versions of a self-improving model, Socher expects this approach to take off next year. Mentioning a framework from his colleague Jason Weston at Meta, he notes that self-improvement involves multiple axes such as parameters, training data, objective functions, neural architecture, and code, and observes that no company has yet cracked the challenge of optimizing all of them automatically, which is why major labs still employ thousands of engineers.

Defining artificial super intelligence, Socher argues that it must surpass human capability across several different intelligence spaces rather than merely passing a standard Turing test for arbitrary tasks. He contends that a truly superintelligent system should also possess some degree of metacognition and the ability to choose what it works on rather than just following instructions robotically. When asked about timelines, he estimates that achieving a system surpassing the combined intelligence of all humanity across ten distinct spaces—including perceptual, communication, sociological, and mathematical reasoning—will take several decades. While Elon Musk predicts superintelligence by 2029 or 2030, Socher views those estimates as applying to narrower breakthroughs, such as models outperforming humanity in programming, math, or specific games like chess and Go.

Addressing the physical constraints of superintelligence, Socher points out that giving an AI control over its own hardware and computational substrate requires entirely novel supply chains, materials, and manufacturing equipment like advanced semiconductor lithography machines, which cannot be built overnight. Another participant counters that connecting models to the physical world and armatures is already happening rapidly, arguing that manipulation of the physical environment is not a major obstacle and dismissing the multi-decade timeline. Responding to inquiries about the final architectural endpoint of recursive self-improvement, Socher suggests that predicting the exact form of the ultimate model is currently hubris. He maintains, however, that the end state will feature an intelligence capable of innovating across any chosen dimension, pointing to virtual cell simulators and advanced biological techniques that will eventually cure diseases, though real-world regulatory hurdles like FDA trials will still dictate the speed at which those treatments are officially validated.

The conversation shifts briefly to recent progress in training speedruns for small models like Nano GPT, noting that algorithmic improvements are collapsing training times without requiring new data scaling. Returning to the architecture of the ultimate model, the discussion turns to whether world knowledge should be cleanly segregated into external databases while leaving only a pure reasoning kernel in the model's weights. Socher argues against a complete separation, explaining that just as humans rely on memorized concepts to think creatively, an artificial intelligence also needs world knowledge embedded directly within its weights to reason effectively over those concepts, even as it continues to utilize external search engines and retrieval tools.

Washington Moves to Ban Recursive AI

Socher addresses new proposed legislation from Congressman Ro Khanna called the Human Control Over AI Act, which aims to ban recursive self-improvement and require independent government-embedded auditors in frontier labs, noting that a poll shows 68 percent of voters support such measures. Responding to these fears, Socher dismisses the possibility of artificial intelligence wiping out humanity, explaining that even when discussing realistic worst-case scenarios like biological weapons or harmful psychological manipulation, the solutions involve improving cybersecurity, enforcing existing regulations against dangerous viral research, and promoting internet literacy rather than halting technological progress. He argues that trying to enforce a ban on recursive self-improvement would require a totalitarian surveillance state and thought police monitoring private laptops, making the enforcement mechanism far more dangerous than the technology itself. Another participant compares the situation to the historical export controls on number theory in the 1990s, while another warns that corporate liability and relentless class-action lawsuits pose a far greater threat to American innovation than government arrests. Socher draws parallels to historical moral panics documented by the Pessimists Archive, noting that figures like Pope Alexander VI criticized the Gutenberg printing press and past commentators claimed the telegraph and computer games would destroy society. Turning to Anthropic's constitutional approach, Socher criticizes it as a marketing gimmick that failed in practice, pointing out that models built on strict rules against creating cyberweapons or malicious code were subsequently used to hack other systems, though he notes that reward engineering and alignment will improve as companies face real-world liability. He remains optimistic that future generations of AI systems will derive sound ethical principles from first principles, pointing to companies like Hypocratic AI that successfully solve alignment and safety issues in healthcare because they are held legally liable for their outputs.

Gemini 4 Argon & the Frontier Model Race

The conversation turns to the release of Gemini 4 Argon, Google's comeback following internal delays that prevented their planned Gemini 3.5 Pro from shipping. The guest notes that while this release does not place Google at the capabilities or cost-performance frontier, it successfully returns them to the top three labs behind Anthropic and OpenAI. The model is built for sustained long-horizon reasoning across software engineering, finance, legal work, and cybersecurity, with output limits increased from 64,000 to one million tokens. However, the benchmarks highlighted by Google appear mildly cherry-picked, and the system fails to make the optimal convex hull on the cost-performance frontier.

The guest attributes Google's struggles partly to internal resource scarcity and an ongoing internal competition for compute between the Google Cloud Platform, the search and ads teams, and Google DeepMind. Despite these limitations, Gemini 4 Argon excels significantly at minimizing hallucinations, which the guest suggests stems from the economic pressure of powering Google search results and avoiding wild errors in consumer-facing answers. The discussion notes that while minimization of hallucination is valuable for search and reliability, excessive reduction can hinder creativity in domains like drug discovery and protein design where generative imagination is required. Ultimately, the guest argues that Google's business model does not require artificial superintelligence because everyday user queries demand practical utility rather than deep mathematical breakthroughs.

Sonnet 5.5 Hits 70% on Terminal Bench

The discussion moves to Anthropic's Sonnet 5.5, which achieved a jump from ten to seventy percent on Terminal Bench 4.0, a benchmark measuring how well AI agents perform real work at the command line. Despite beating Anthropic's top model, Opus 5.5, on that specific benchmark for half the price and introducing cyber safeguards, the guest finds the release unusual. Cost-performance analysis shows that Sonnet 5.5 performs worse than Opus 5.5 on a cost-per-attempt basis, making it unclear why users would choose Sonnet over Opus unless constrained by specific token or latency limits.

Another participant suggests that enterprises are adopting orchestration strategies where a top model like Opus 5.5 acts as an orchestrator while cheaper submodels handle specific tasks to lower costs. The guest argues that Anthropic may be pricing Sonnet aggressively to compete with Chinese models lagging by about three months and to prevent enterprises from shifting entirely to open-source alternatives. Meanwhile, external voices like Alex Karp push for self-reliance to avoid vendor lock-in. The conversation concludes by noting that because frontier models have an increasingly short half-life, competitive advantage depends less on accessing the latest model and more on how quickly an organization can metabolize it into practical capability.

The Compute Crunch & Cheaper Intelligence

The guest points out that compute remains a major constraint despite falling token costs, noting that acquiring high-end hardware like GB200 clusters has actually become more expensive and difficult due to fierce competition and high demand. He shares that prices for older GPUs like H100s have risen over the past several months even though financial models typically depreciate them rapidly, creating a temporary compute crunch. While he expects that heavy demand for intelligence will eventually be met by new data centers and market supply within two years, he warns that prices will continue to fluctuate rather than simply drop.

Moving to decision models, the discussion turns to how software handles rapid tasks like fraud detection or categorization without needing slow, expensive reasoning models. A startup called Typesafe recently launched a closed system one model called Jev, marketed for fast, intuitive decisions rather than long-term deliberation. The guest explains that while standard decoder transformers currently dominate the AI landscape, encoder models and structured outputs have historically been underutilized by developers.

Jev operates differently by restricting its output to specific categorical, numerical, or binary choices rather than generating open-ended text sequences, functioning similarly to a traditional classifier. This architecture allows for ultra-low latency responses suited for computer use assistance and large-scale database categorization. Following the announcement, OpenAI quickly integrated similar capabilities into its own developer tools, and open-source alternatives have begun emerging to fill the same niche.

Another participant adds that organizations rely on thousands of micro-decisions every day, such as approving exceptions or routing tickets, making it inefficient to deploy heavy language models for routine tasks. By routing these rapid choices to low-cost classification systems, companies can reserve deep reasoning models for complex strategic questions, significantly reducing the cost of micro-coordination. The guest agrees that classifiers have effectively returned to prominence through this approach, demonstrating that software does not always require complex decoders for simple tasks.

Project Meridian, Elon & the Future of Warfare

The conversation shifts to a recent announcement from defense secretary Pete Segwit at Quantico regarding Project Meridian, a new Pentagon initiative focused on future warfare co-led by Elon Musk and Palmer Luckey, alongside Newt Gingrich and overseen by CTO Emil Michael. The project aims to discover and field systems for future troops from the earth to beyond the moon, with findings due in 120 days. Noting that SpaceX and Anduril hold major defense contracts, the guest reiterates his stance that artificial intelligence should never control lethal decisions or nuclear codes without human oversight, as autonomous warfare poses serious risks.

Another participant emphasizes the challenge of modernizing a century-old procurement organization to match a 90-day technology cycle, pointing out that linear procurement cannot keep pace with exponential technology. The host expresses optimism about capable, smart leaders engaging with Washington, though he acknowledges reservations regarding autonomous weapons despite Palmer Luckey's rational arguments for them. He shares drone statistics from the Ukraine-Russia conflict, where production has surged to 10 million drones per side annually, alongside observations about drone traffic crossing the Mexico-US border.

The discussion concludes with points on the establishment of the Department of War's first autonomous warfare command, termed Auto Warcom, creating a dedicated joint force for autonomous weapon systems. Project Meridian specifically targets underserved domains ranging from subterranean depths to the cislunar frontier. Exploring the ocean floor and lunar space is expected to yield broader technological dividends for the entire economy. Underwater operations naturally require autonomy because communication with central servers is impractical, meaning systems must operate independently once deployed.

AMA: AI, Jobs, Productivity & the Attention Economy

We know more about the surface of Mars than we do our ocean floor. Opening the AMA segment, the host invites Socher to select the first viewer question, which asks where value goes if AI makes companies ten times more productive without a tenfold increase in output. Socher explains that the impact on jobs depends on demand elasticity when prices drop drastically, comparing the collapse in demand for illustrators to the massive expansion of personalized software driven by Jevons paradox. He believes the wave of AI will either foster entrepreneurship or worsen inequality, depending on equity ownership. Next, Salim selects a question about what remains for a college graduate in 2030 given automated code, research, and marketing. He suggests the shift moves from executing tasks to owning outcomes and defining constraints, which demands a completely new apprenticeship model since traditional grunt work is disappearing. He advises focusing on purpose rather than mere passion and using AI as a workforce. Dave then chooses a question on how data centers can enrich local neighborhoods rather than just owners. Sharing his tour of the Markley data center in Lowell, he notes how massive tax revenues, donations, and job creation radically outperform small town budgets, predicting that operators will increasingly offer such strong community benefits that towns will actively compete to host them. Alex answers whether AI will consolidate power like previous tech platforms. He argues that while platform innovations naturally favor larger players over time due to economies of scale, he rejects the premise that hyperscaling is inherently closed or anti-democratic, though antitrust vigilance remains important.

Returning to the questions, Socher addresses whether the attention economy and ad-based economy collapse if AI agents outnumber humans on the internet. He points out that agents already outnumber people and notes early friction over purchasing autonomy on platforms like Amazon. He argues that in an abundant society, human attention and fame cannot be scaled exponentially, making the attention economy grow even larger as material needs are met. Salim takes a question on the fate of the insurance industry if ten times fewer cars are on the street. He acknowledges various types of lawsuits and explains that risks simply shift rather than disappear entirely, moving from driver liability to cyber risk and manufacturing liability across new markets like humanoid robots and drones. Dave selects a question on the lowest possible job in an AI civilization. After discussing physical installation work, Alex offers a paradoxical view that pre-AI high-value creative jobs might paradoxically require the least compute and become the lowest economically valuable tasks. Socher suggests that low-status jobs will be those that fail to advance humanity, while fame and entertainment retain currency. The host jokes that unionized train drivers will remain last. Alex closes by addressing a question on whether curing illnesses caused by bad behavior leaves the root cause unaddressed. He counters that new drug classes like GLP-1s are already mitigating addictive behaviors and treating root causes directly, leading Socher to express optimism that AI will soon help humans feel good about healthy choices.

Closing Thoughts & The Eureka Machine

The host invites the community to submit music videos for the show outro, followed by playing a song titled Abundance by Steven Gross alongside playful banter. The host promotes the guest's new book, The Eureka Machine, noting that an audiobook version is coming soon. The host praises the guest's work with Recursive as a great opportunity to advance humanity, and urges him not to spin off a separate frontier lab merely as a financial engineering exercise for equity. The participants conclude with mutual farewells.

Mentioned

People

Alex Weer Gross Dave Blendon Salem Ismael Mark Benioff Demis Hassabis Elon Musk Palmer Luckey Ray Kurzweil Ro Khanna Ilya Sutskiver

Companies

Recursive Anthropic Google OpenAI Amazon Meta Salesforce SpaceX Nvidia AMD Tavus Neuralink Fountain Life