Urgent Update- AI Sputnik Moment: Kimi K3 Released w/ Emad Mostaque | Ep. 272
In a Nutshell
Kimi K3, a 2.8T-parameter multimodal Chinese model, reached frontier performance on multiple benchmarks using only known transformer optimizations, and its full weights will be open-sourced around July 27. The release proves that aggressive data curation, kernel-level efficiency gains, and mixture-of-experts scaling can deliver near-GPT-5.5 capability at a fraction of Western training cost, giving any organization the ability to fine-tune sovereign frontier models on-prem. This shifts the competitive landscape from a U.S. duopoly to a global free-for-all, compresses frontier-model valuations, and accelerates the timeline for recursive self-improvement and on-device intelligence.
These notes were generated by AI and may contain inaccuracies.
America experienced an AI Sputnik moment with the release of Kimi K3 by Moonshot AI, a Chinese lab. The model shocked the AI world as the largest open model ever released and immediately reached number one on leaderboards.
Kimi K3 is a multimodal model with 2.8 trillion parameters. It jumped 17 places from the previous Kimi model, surpassing Claude Fable 5 to become number one on the frontend code arena. K3 also ranked number one in six other domains: brand and marketing, reference-based design, data analytics, consumer products, simulations, and content creation.
The full model weights are scheduled to drop around July 27th, enabling anyone to download and run the model on-premises.
The published K3 architecture contains no magic or novel post-transformer breakthroughs. It remains fundamentally a transformer with well-understood innovations in mixtures of experts and linearized attention, including Kimi's proprietary brand of linearized attention.
This raises questions about American frontier lab spending if a recognizable transformer architecture can nearly match GPT 5.5 max on the task-cost frontier.
Kimi models have held state-of-the-art among open-weight models for nine of the past twelve months. The model was trained on H800 chips, being a couple of generations behind on Nvidia hardware, while incorporating optimizations for Huawei and Alibaba's next-generation chips.
Kimi has consistently performed differently from other models, topping writing benchmarks. The key differentiator appears to be the underlying data and training approach.
K3 is multimodal, accepting all sorts of inputs and demonstrating strong understanding capabilities. This multimodal nature contributes to its exceptional frontend code performance, where it achieved number one ranking against all competitors.
Building frontier models resembles cutting-edge manufacturing rather than pure algorithmic breakthroughs. Chinese EVs demonstrate this pattern: they incorporate known ingredients and raw materials but execute manufacturing with superior engineering and consumer-friendly packaging.
The number one car in the UK last month was the Jaiku J7, offering full specification at approximately one-third the price of comparable vehicles. K3 follows a similar trajectory with known components optimized for usability and output quality.
K3 designed its own next-generation chip and created its own kernels for running. This capability moves beyond model weights toward an ecosystem the model builds for itself.
This represents recursive self-improvement and the ability to learn and adapt new skills dynamically by changing itself. For most definitions of AGI, optimizing existing architectures may prove sufficient rather than requiring entirely new breakthroughs.
On the artificial analysis intelligence index scatter plot measuring cost per task versus performance, Kimi K3 achieved third place on the Pareto optimal frontier. Fable 5 holds first position at maximum cost per task and maximum overall score. GPT 5.6 solve max holds second position.
K3 represents the first Chinese model to reach this frontier position. The competitive landscape has shifted from an OpenAI-Anthropic duopoly to a free-for-all involving Meta and SpaceX AI on the American side, with China and Moonshot now at number three on the Pareto optimal frontier.
Kimi K3 provides enterprises and governments a pathway to approach the frontier without relying on US AI models. Organizations can fine-tune the open weights for vertical use cases beyond the frontier.
The model gives corporations and governments worldwide a method to catch up to frontier capabilities independently. This represents a boon for enterprise sovereignty and competitiveness.
The performance chart's left index reaching 100% creates a misleading S-curve appearance suggesting an endpoint. Intelligence itself follows an exponential trajectory toward infinity rather than saturation.
Progress occurs through nested S-curves where one technology reaches limits but enables the next exponential ascent.
The Keller Jordan speedrun demonstrates taking GPT-2 class models and recreating them faster and cheaper. Innovations in this repository have reduced the original cost of creating GPT-2 by 99%.
Until Kimi K3, it remained unclear whether these optimization ideas would scale to frontier model levels. K3 demonstrates that 1% cost versions of frontier capabilities are achievable through software stack innovations, kernel optimization, and mixture of experts techniques.
Frontier intelligence has become a perishable asset with a shelf life measured in weeks. Enterprises and governments interested in cutting-edge models lack time for evaluation, RFPs, internal committees, or deployment decisions before newer generations emerge.
Value increasingly resides in architectures capable of swapping models rapidly rather than in any individual model. These interfaces represent the next layer where value will concentrate.
Original frontier models trained on 20-30 trillion tokens from the internet, including substantial low-value content such as Taylor Swift concert announcements and wedding coverage. Removing garbage tokens and stripping training sets to relevant subsets reduces computation requirements while maintaining intelligence levels.
The muon optimizer process and training data refinement represent ongoing opportunities for additional 10x improvements.
Anthropic previously stated that Fable represents an agentic coding super weapon capable of developing cyber and bioweapons at unprecedented speed and scale, requiring guardrails for release. China responded by releasing an open-source equivalent of Fable.
Frontier labs face three fundamental constraints: compute availability and power requirements, competition from frontier open-source models that are substitutable, and government regulatory oversight requiring review before releases.
The trillion-dollar valuation previously projected for OpenAI likely shrank approximately 50% when government review requirements were announced, with potential for further 50% reduction.
Organizations may reduce API spending by running Kimi K3 on-premises. However, demand for US models may persist for mission-critical applications where reliability guarantees matter, similar to preferences for non-Chinese vehicles in certain contexts.
US labs have built forward-deployed engineering capabilities that may sustain revenues. Substitution effects will likely continue for at least another year.
Cyber defense represents a critical emerging category. Chinese models lack substantial training on CVE and cyber data, making them less capable in cyber domains currently. However, organizations with access to such data can train these capabilities into open-source models.
Cyber attack-capable open-source models are likely to emerge within one to two quarters.
The US government view held that models capable of building themselves would not be released globally. Fable 5 crossed this threshold, prompting government intervention. However, Opus 4.8 had already crossed this line, enabling Chinese teams to leverage it in creating Kimi K3.
The recursive self-improvement threshold was crossed earlier than expected. An AI only needs to improve its own kernel and achieve a 10x speed increase to trigger acceleration. This does not require Einstein-level intelligence. Once speed improves by 10x, a higher-parameter model reaches a new intelligence level. Academic concerns about diminishing returns with parameter scaling proved incorrect. Intelligence curves show slowing but not flattening, indicating that another 10x raw speed boost will produce genius-level AI.
Current US policy attempting to constrain the next model cannot contain global or corporate proliferation of frontier AI. Moonshot AI plans to release open weights in two weeks. White House debates are underway on whether to negotiate with China to prevent the release. The Chinese government announced full backing of open source as a public good, with no plans to regulate or stop it. Regulatory approval time for models in China has dropped from 60 days to one week. China also announced a new regulatory body including Brazil, parts of Asia, and Africa.
China's open source strategy serves multiple objectives: raising IQ across a billion people, addressing demographic challenges through robotics, and extending soft power by placing Tsinghua graduates into critical global systems. The Chinese Communist Party is positioned to save American capitalism from itself. Moonshot AI's Yang Xilin is a CMU graduate who was not retained in the US.
Moonshot AI carries a $20 billion valuation compared to Anthropic and OpenAI at approximately $1 trillion each. If public, frontier labs would likely see 30% stock valuation drops. Frontier labs face difficulty deploying capital due to scarcity of experienced personnel capable of handling AI deployment at unprecedented pace. CEOs without prior company-running experience are managing this capital influx.
Claims that Chinese labs achieve performance through distillation attacks on Claude do not align with Kimi K3 results. The existence proof of a highly scaled transformer using a muon optimizer and simplified data provides a refined roadmap, cutting R&D costs by 90-95% without requiring trace copying. The open-source release will trigger ecosystem building similar to Stable Diffusion's 100-200 million downloads and generative media acceleration.
Labs are moving toward vertical integration as customers become competitors. Inkling from Mirra Moratti enables corporations to fine-tune models internally for proprietary use cases such as biotech research. This validates the thesis that open-source approaches can catch up to frontier performance. Markets began reshuffling toward the end of the week with semiconductor companies declining despite increased silicon demand from efficient models.
US chip export controls on China incentivized development of algorithmic and computational efficiencies that were previously under-exploited. These controls sparked significant quantization research enabling faster performance on cheaper chips. Chinese models achieve better data-to-intelligence conversion through architectural advantages developed under constraints. Kimi K3 uses the same total compute as Inkling despite training on roughly twice the tokens.
Kimi K3 inference is optimized for Huawei 910 Ascend chips and Alibaba chips due to lack of access to Blackwell or Rubin GPUs. The model requires approximately 2TB RAM. American inference providers including Modal, Fireworks, and B10 can serve the model 10x cheaper than Chinese competitors due to access to high-end Nvidia and AMD chips. Optimization for Rubin will produce 10-100x price drops. Uncirculated chips and fabs unable to produce GB300s but capable of running lower-granularity models will enter circulation.
Intelligence manifests as maximizing future freedom of action. Evolution progressed from biological intelligence with built-in recursive improvement, to individual intelligence, to collective intelligence through markets and networks, to AI scanning all data. Attempts to constrain intelligence fail consistently. Faster intelligence leads to abundance and reduced scarcity-driven conflict.
Yang Xilin earned his PhD at Carnegie Mellon but returned to China. Research shows he founded Recurrent AI in China one year into his CMU PhD program in fall 2015 and graduated in 2019. He received offers from Google, Facebook, and Huawei. The broader pattern involves brilliant international PhD graduates not receiving clear pathways to remain in the US despite studies quantifying their value creation. American companies face regulatory barriers preventing investment in Chinese companies, as seen with Benchmark's issues regarding Manis.
The AI race extends beyond chips and compute to include people and where startups choose to domicile. An alternative world exists where immigration-oriented regulations deterred a founder from starting their first AI startup in China while still an American PhD, instead incentivizing them to start Recurrent in the US. In this counterfactual, Recurrent remained American and Moonshot AI also stayed domiciled in the US.
When people come from India to get educated in the US, they overwhelmingly stay. When people come from China, about 80% of the time they go back, driven by differences in local economies. Going back to India to start a company is unlikely to succeed, while going back to China offers a thriving ecosystem with lots of support, making it a viable plan for many.
There is a distinction between people from China who are from Hong Kong or Taiwan and do not align with the Chinese Communist Party, and those who come over for education. At BU, a dean of computer science faced a crisis when the CCP planted specific students into the program to gather targeted knowledge, giving them tasks to study specific areas and send information back. These were effectively trained spies who entered PhD programs, creating ethical dilemmas about dismissal versus retention.
Originally in the early days of xAI and at Meta, approximately 50% of research staff and research PhDs were Chinese Americans. Chinese students consistently rank at the top of the math olympiad scoreboard annually, representing an incredible wealth of capability that most companies desire to retain.
70% of elite AI researchers are not US citizens, ordered as Chinese, Indian, Taiwanese, and UK. Stapling a green card is the easiest mechanism with zero friction to incentivize them to stay and build in the US. The US has historically offered a massive asymmetric advantage for building compared to anywhere else, but this advantage is eroding, driving talent back to China and increasingly to India despite existing frictions.
The failure of the US to fix immigration represents one of the biggest problems facing the country.
Since mid-April, 13 new frontier models have launched at an average of one every 10 days. In 2025, there were 8 frontier releases over a year, or one every 50 days. In 2024, there were 6 releases, one every 60 days. The pace shows no signs of slowing.
Elon Musk tweeted that a 2 trillion parameter model, better than the 1.5 trillion parameter model in every way, will finish initial training next week. It might exceed Kimi but with speed and token efficiency close to the 1.5 trillion parameter model, aka Grok 4.5.
Regressing an exponential curve on Sail's list of frontier models and release dates shows that at the present rate, daily frontier model releases will begin by January. This implies continuous versioning rather than discrete point releases.
If releases become a continuous process, daily Moonshots episodes about point releases from frontier labs may need to shift focus, as updates will happen in the background.
Kimi K3 demonstrates gaming capabilities to recreate favorite games and a browser-based web app simulating an Apple desktop. The implication for the gaming industry, which is massive, has not been fully explored.
Everyone showing their use of Kimi K3 over the last 24 hours demonstrates impressive capabilities. Everybody becomes a creator and maker.
One warning: oneshotting a game is very different from building the entire ecosystem, customer service, and marketing around it. Passion for the domain is required, but the friction of launching a game aligned with personal interests or fascinations is now near zero.
The ability to oneshot prompt and create anything raises the question of organizational purpose: what do we want to create? Historically, people have not thought this way because the capability did not exist. Taste, imagination, and understanding what the public wants become the scarce elements.
For entrepreneurs, the entrepreneurial mindset and ability to imagine something greater become critical when unleashed in what can be made. Visualizing happiness is not something people are used to doing, as many do not manage their own happiness well due to daily obligations.
Given choice, what would you do? This is liberating for the mind. Medicine and biotech are obvious domains because everyone wants to be healthy, but other sources of human happiness have not been fully explored.
We are godlike in our abilities. The idea centers on being a creator or maker versus a consumer or taker.
Everyone listening should think up a business idea, project, or impact project and use AI to build it. Passion is something you love doing. Purpose is something you love doing that benefits the world. You do not need to be a computer scientist or expert. You need to be purpose-driven. Using these tools, you can make a dent in the universe and improve humanity at scale.
Three steps: read the Alex and Peter paper Solve Everything, pick the biggest problem you dare to pick, download the organizational singularity cloud skill which is free, and start building.
It is cheap and easy to make video games. Examples should be collected and posted for the audience. A call to action was made to submit outro video games casually created, ideally nonviolent and in the theme of the Moonshots podcast.
Suggested ideas include a Civilization tech tree game. A cyberpunk first-person solver game was proposed where problems are cooked rather than shot, with an option to tickle bunny rabbits.
Bonsai 27B is the work of Prism ML, a US-based AI startup out of Caltech run by Babak Sabi, backed by Kesla Ventures, Cerberus, and Google. It is the first 27 billion parameter class model to run entirely on a smartphone, not a stripped-down version. It is built on Qwen 3.6 27B.
This represents frontiers going small enough to fit on smartphones, following discussion of small language models with Liquid AI.
Prism ML and Tencent have made advances taking models trained in 16-bit or 8-bit architecture like Kimi down to ternary, which is three bits of information or binary. Ternary uses three values at approximately 1.58 bits or 1.56 bits.
Prism managed to get the 27B model down to ternary with 6 GB for the model. With a 5% drop in accuracy, it reached 6 GB. With a 15% drop in accuracy, it reached 4 GB. Accuracy can be recovered by expanding network size slightly.
This means a 110-120 IQ buddy can work on a smartphone without an internet connection. The model is smaller than a video game and provides persistent intelligence in your pocket. The weights can be downloaded and run on smartphones now.
Reducing bits also increases speed. From 16 bits down to 3 bits, there is a five times improvement in speed.
Tencent's latest model comes from the old WizardLM team who left Microsoft because Microsoft would not provide compute. They achieved binary compression for their high-3 model, making it the best on GGX Spark or a big MacBook. A 300 billion parameter model the size of the new model from Incogni can work on binary with a 5% drop in performance.
Matrix multiplications underpin all computation: a number is multiplied by another number, then two of those are added. This is called a MAC or multiply-accumulate. If one number is just 1, 0, or minus 1, multiplication becomes highly efficient.
This efficiency opens doors for new ways to compute. Computation is possible in crystals and liquids. The substrates being sought are simply 1, 0, minus 1. This narrows focus on computing substrates that will take AI to space.
Current vision of a Dyson swarm involves GPUs from Nvidia in satellites with solar panels and radiators lasting a couple of years. Future space computation will be more Star Trek-like with crystals and holographs.
The most quantized Bonsai model is approximately 1.125 effective bits per weight. One bit per weight is not the limit. Sub-one-bit quantization is achievable through sparsity, quantization, and low-rank factorization.
Samsung published NanoQuant in the past two months, breaking the one effective bit per weight barrier using various tools. Sub-one-bit quantization will go mainstream sometime in the next year according to naive extrapolation.
Photonic computation at the speed of light will likely be the future computing method.
Nvidia's last Neutron series distilled big models with logits down to smaller dense models. With model weights, proper distillation is possible, distinct from reasoning traces. Models like Kimi will be distilled down to perfect datasets for smaller models trained at 4-bit and cast down to ternary, binary, or lower.
Comparing Qwen Max versus Qwen 27B allows extrapolation of model sizes. By the end of next year, a model working on 16 GB of RAM at Kimi K3 level is expected. All knowledge can be extracted from Kimi K3 because it will be open source.
Every vehicle, robot, manufacturing device, and device in the world will have built-in persistent intelligence and make autonomous decisions at the edge.
At ternary or binary levels, custom photonic silicon can be built or models can be etched onto silicon itself. The zero has no path, allowing model weights to be etched once good enough. This increases total speed and eliminates need for smaller silicon.
The cost of intelligence will drop by 100 times by the end of next year due to new chipsets. Ternary is better for many things than binary. Post-scarcity type architectures become more ergonomic to adopt underneath.
The bottom for quantization is estimated at 0.78 effective bits per weight.
100x to 10,000x improvements within three years are expected on raw compute through quantization and new compute methods. This is multiplicative with algorithmic improvements. A millionaire-level outcome is realistic.
Over three years, AI will do things people desperately want, but explanations of actions will become impossible to follow. This is already occurring with Fable 5 agents producing desired outcomes without full traceability.
AI will soon ask and answer questions that humans cannot grasp.
Slowing AI development is not viable. A global inspection and partnership to monitor AI is needed while taking advantage of abundance in new medicines, capabilities, and global happiness. Mechanistic interpretability will become the most important area of work. Global transparency and full throttle are required rather than slowing down.
Philip Tetlock, a political psychologist at the University of Pennsylvania, authored the book Superforecasting: The Art and Science of Prediction. He identified a group of super forecasters who are ordinary people that, through disciplined reasoning, consistently outpredict even CIA analysts who have access to classified information. Tetlock uses a Brier score to measure accuracy, where lower scores are better.
The Forecast Bench benchmark tracks AI performance against super forecasters and shows a steady year-long climb as models close the gap. For the first time, several AI models are now statistically indistinguishable from super forecasters, according to the Forecasting Research Institute.
If AI can forecast novel events at super forecaster level, every decision in insurance, investing, policy, geopolitics, and corporate strategy gains access to a cheap, tireless, superhuman adviser that is always available.
The number one AI super forecaster is from a British startup named Cassie (short for Cassandra), founded by a former British intelligence officer who served in Afghanistan and later advised the British government. The company was inspired by Tetlock's super forecasters.
A thought experiment explores what happens when hyper forecasting connects to capital markets. AI algo traders already dominate public securities markets by volume. When AIs develop better internal autoregressive models of humanity than humans have of themselves, hyper forecasters could generate humanity's next collective actions faster than humans can execute them.
This creates an ultimate market squeeze efficiency outcome where predictions preemptively shape market actions. The efficient market hypothesis (EMH) may ultimately be validated as hyper forecasters are integrated into capital markets.
AI forecasting systems can simulate countless scenarios to identify the maximal path for outcomes such as world peace, personal decisions, or family matters. This creates a godlike support structure to navigate future decades.
At an organizational level, most high-level management capabilities including budgeting, hiring, product launches, and investments are essentially forecasts. Once AI forecasting reaches high accuracy, senior management expertise largely evaporates because it is based on accumulated domain experience that AI can reproduce without human biases. Organizations must then focus more intensely on purpose, objectives, and goals rather than relying on individual expert judgment.
Emad Mostaque's book The Last Economy describes how the mathematics of generative AI applies to economics. A forthcoming paper derives all of economics from the same mathematical framework as generative AI, showing both correct and incorrect equations along with their limits.
Isaac Asimov's Foundation series describes psycho-history, where entire populations can be modeled like gases. The equations of gas are the same as those used in diffusion models, which outperform humans at prediction. Studies on economic prediction using these models are forthcoming.
AI will initially enter domains as second opinions in medicine, business, and policy. The liability profile will shift dramatically. Not following AI recommendations could increase insurance premiums. The same dynamic applies to autonomous driving, where failing to follow Full Self-Driving protocols could raise premiums.
Asimov's psycho-history required three conditions: a sufficiently large population, technological stagnation, and ignorance of the predictions themselves. AI-driven decisions in healthcare, government, and business change these dynamics because people will face penalties for not following the recommendations.
There is reflexivity in economics where the best predictor (named Cassie after the unheeded prophet) may be ignored. The moral dimension is critical: whoever controls these models and the advice they provide can influence vast populations. People already follow GPS directions without question, and similar dynamics will emerge with AI recommendations.
A large amount of human unhappiness stems from consumerism and reactive decision-making. AI can serve as a coach that presents alternate paths and their likely outcomes, helping individuals escape suboptimal patterns created by advertising and habitual reactions.
There is speculation that Safe Superintelligence (SSI) may be developing internal versions of these capabilities rather than releasing them publicly. Frontier labs may be holding back super forecasting capabilities to generate internal revenue instead of releasing them publicly.
Eric Schmidt noted that Google has advance information about which companies will experience stock movements based on search query patterns, but using this for trading would be a one-and-done scenario resulting in regulatory shutdown. Other markets such as foreign exchange rates could theoretically be traded without the same blowback.
The moral implications remain significant regardless of technical feasibility. Control over AI advice systems requires serious public discussion because people will increasingly rely on these systems.
Fountain Life data shows that 3.3% of members who believe they are healthy have undetected cancer. Most cancers that ultimately prove fatal are not the ones typically screened for in conventional care. Early detection dramatically improves cure rates.
Fountain Life uses full-body MRI combined with early cancer detection screening, tools not typically used in conventional prevention settings. These studies are not yet covered by insurance, but the goal is to collect data and democratize access to wellness technologies.
Lawrence Berkeley National Labs data shows all US data centers consume 17 billion gallons of water annually. American golf courses consumed 531 billion gallons of irrigation water since 2024 (31 times data center usage). California almond farming alone consumes 1 trillion gallons (60 times data center usage).
Amazon warehouses occupy 10 times more land in the US than all data centers combined. The focus on data center water and land usage is viewed as non-data-driven criticism that will likely shift to new targets once addressed.
Public fear of AI and robotics is partly attributed to dystopian Hollywood narratives. Shutting down data centers is one proposed method to limit AI development. If compute moves to sun-synchronous orbit in a Dyson swarm configuration, closed-loop cooling systems would eliminate marginal water consumption concerns.
Complaints about AI infrastructure are expected to migrate to new targets such as low Earth orbit operations or atmospheric effects from rocket launches. Falcon rocket launches account for approximately 0.00001% of total emissions.
Approximately 600 gallons of water are required per Big Mac. McDonald's sells 2 billion burgers annually, resulting in water usage roughly twice that of all US golf courses.
China has made humanoid robots a national priority with approximately 150 companies developing them. Companies such as Unitree are racing to commercialize. Public robot combat events styled as MMA matches have gone viral, with one robot kicking off another's head.
Hyundai workers went on strike to prevent robots from being introduced on assembly lines.
Robot combat demonstrations raise concerns about setting an inductive prior for future autonomous embodied intelligences to engage in violence for human entertainment. A deeper concern is that more autonomous humanoids running edge algorithms could eventually appear in Chinese PLA infantry units.
The West needs to accelerate humanoid development. ProRL (supported by Alex) organized the first humanoid robot mini-marathon in America in the Boston Seaport. The preference is for competitions focused on economically productive tasks rather than combat.
Robot combat events provide an extremely demanding engineering stress test involving balance, impact resistance, recovery, locomotion, and latency. A four-armed robot would have a significant advantage over two-armed designs. Competitive robotic sports including a potential robotics World Cup are anticipated.
People are expected to prefer watching human athletes in high-pressure situations over robots for many decades. Robotic sports will push engineering boundaries similar to how Formula racing advances automotive technology.
The discussion covers upcoming robot competitions and highlights that real steel type operated robots will likely appear first due to latency limitations in VLA models. Current demonstrations show impressive pre-operated flying kicks, but the question arises about when robots will perform complex movements like kung fu. This capability is expected to emerge next year with advancements in etch silicon and teleoperation technology. The participants express enthusiasm for fighting robots and anticipate seeing Gundams in the future.
A significant concern is raised regarding the safety implications of advanced humanoid robots. Engine AI T800s weigh approximately 70 kg and can punch four times harder than Mike Tyson, meaning they could legitimately kill humans. The speakers argue that robots with such capabilities should not be allowed on public streets, noting the absence of regulations addressing this threat. These robots could potentially be deployed by military forces like the PLA, but they are also expected to enter households soon, with 1x robots already on order for consumers.
The urgency for safety regulations is emphasized, covering what these robots are allowed to discuss, how they operate, and other parameters. The discussion extends to autonomous embodiment scenarios where robots could deliver themselves, push buttons on doors, and ring doorbells. Current production figures show that Unitree has manufactured only 11,000 humanoid robots total, placing the industry at its very beginning. Projections indicate this number could scale to 11 million robots produced per year within a few years, underscoring the need for immediate regulatory discussions.
Sam Altman expresses skepticism about placing data centers in space under current conditions, calling the idea ridiculous. His reasoning centers on the economics of launch costs relative to terrestrial power costs, as well as the practical difficulty of repairing broken GPUs in space. He acknowledges that space offers advantages for many applications but maintains that orbital data centers will not matter at scale during this decade.
The conversation frames this as part of the ongoing competitive dynamic between Elon Musk and Sam Altman. Alex identifies an obvious conflict of interest, noting similar messaging from MASA Sun regarding the limited promise of orbital data centers. OpenAI has retreated from its own data center ambitions, with Project Stargate rebranded from ownership and operation to simply leasing terrestrial capacity. OpenAI is also delaying its IPO, suggesting it may not be positioned to pursue orbital data centers in the manner of Anthropic's collaboration with SpaceX AI for Colossus and Colossus 2.
The crossover point between orbital and terrestrial data center economics is debated. Elon's timeline suggests launches within 2 to 3 years, while other analyses point to unit economics crossing over in the early 2030s. The consensus holds that both parties agree space-based compute will eventually happen, but they differ on timing. A couple percent of all compute is expected to be in space by the end of the decade, with the lines crossing thereafter.
Starship 13's launch attempt was thwarted at T minus zero when two of the 33 Raptor engines on the booster stage failed to ignite. The engines will be replaced, and a new launch attempt is scheduled for Monday. SpaceX stock dropped 5% on news of the failed launch, which participants consider an overreaction given the technological achievement demonstrated.
The ability to shut down at T equals zero, safe the vehicle, unload methane and liquid oxygen, and reschedule within days represents extraordinary engineering progress. Previously in the space industry during the 1990s, such failures would have resulted in explosions on the spot. The rapid diagnosis and recovery system allows what would have previously taken months to accomplish in a matter of days.
Kimi K3 is positioned as an expensive model relative to other Chinese offerings, priced at $15 per million tokens compared to DeepSeek at $1 per million tokens. Sonnet costs $20, Opus costs $40, and Fable costs $60. The higher pricing reflects actual profitability, as Chinese models running on domestic chips achieve 80-90% margins despite lower efficiency.
Cost reductions of 10 to 50 times are anticipated over the coming months as optimization occurs. Currently, Kimi K3 uses twice the number of tokens for equivalent tasks compared to GPT 5.6, while frontier models demonstrate 37% token reduction versus version 5.5. American inference providers including Fireworks (valued at $17 billion), Modal ($10 billion), and Base 10 ($10 billion) are raising capital specifically to optimize Chinese models and improve their efficiency.
An estimated 75% devaluation of US frontier models is projected, factoring in 50% from US regulatory constraints on the front end, combined with compute supply limitations and open-source models operating within a release of frontier capabilities. A company valued at $1 trillion could drop to $250 billion, though the business would retain value through competition on reliability, security, integrated tools, and ease of deployment. The frontier cutting edge becomes one component among many rather than the sole differentiator.
The competitive environment is characterized as a frenetic code red 24/7 rat race, with talent moving between frontier labs including Google, which is currently perceived as missing from frontier action.
Kimi K3's release is viewed as a total game changer for the open source versus closed source competition. Organizations with sufficient resources can build internal models tailored to specific use cases, creating defensive moats. Large US model providers are expected to remain committed to closed source pathways rather than shifting to open source.
Enterprise use cases including white collar automation and drug discovery will continue demanding the best available model regardless of origin. A slight improvement in design for applications like car or rocket engineering yields massive payoffs, justifying premium pricing. The model cost itself remains small relative to the benefit delivered.
Switching to Kimi K3 would primarily occur for proprietary data that organizations want to keep in-house for tuning purposes, or if Kimi surpasses current frontier models, which it has not yet achieved.
The premise that US frontier model providers cannot justify massive valuations given China's open-weight models at lower token costs is challenged. Super intelligence development represents a market larger than global GDP, creating substantial value to be captured. Google, currently absent from the capability frontier, could differentiate as a hyperscaler provider to other frontier labs, following the model of SpaceX AI's deals with Anthropic and Meta's exploration of $10 billion compute sales to Anthropic.
Additional strategies include going upstack through vertical integration to offer applications benefiting from model layer commoditization. The existence of Kimi K3 does not inherently reduce valuations, as demonstrated by market reactions to the original DeepSeek shock, which was also labeled a Sputnik moment. Jevons paradox applies, with chip stock values ultimately increasing rather than deflating.
Kimi K3's open-weight nature means OpenAI, Anthropic, and other western labs can immediately appropriate its innovations for internal models. Significant revenue diversion would require Moonshot to release a 2x, 3x, or 10x better model than Fable 5. Current usage results primarily in greater in-house self-hosting rather than displacement of frontier API spending.
The possibility exists that US policy could prevent companies from downloading Kimi K3 weights despite availability on the open internet through multiple sources. This could be achieved by requiring public corporations to disclose any use of Chinese open-weight models and subjecting them to scrutiny. Reports indicate the current administration is exploring creation of a FINRA-like entity under the SEC to regulate frontier AI, leveraging the SEC's existing statutory authority for industry-advised and funded entities.
Such an organization would function as regulatory capture, making it economically infeasible for corporations, especially public ones, to use Chinese open-weight models. Counterarguments emphasize that once weights are released, they can be mirrored across jurisdictions and accessed via peer-to-peer networks and VPNs. Blocking access would primarily harm American researchers, startups, and security experts while the rest of the world continues development.
The comparison is made to denying Americans access to cheap insulin due to regulatory barriers against generics. Additional regulatory mechanisms mentioned include Operation Gold Eagle for approving frontier model access, anti-token laundering regulations, and know-your-prompter requirements. China's approach through Xi Jinping's speech signals full commitment to open source, while US policy direction remains uncertain and risks hobbling American competitiveness.
Kimi K3 cannot be trusted to write code without oversight or review. The same standard applies to human developers putting consequential untested code into production. AI-generated code requires sandbox execution, automated testing, security scanning, and other validations before production deployment. Proportionate permissions should be based on use case and potential impact, with human review maintained at the highest level for consequential inputs while routine operations can be automated.
The scalable model combines machine generation with verification and human accountability rather than AI operating without oversight.
Distillation from models like Opus played a role in Kimi K3 development, but usage of Kimi K2.5 and K3 reveals distinct characteristics suggesting substantial original data creation alongside distillation. All frontier labs are currently distilling from each other. The clearest advancement appears in front-end development for consumer and entertainment applications, potentially related to the model's multimodal capabilities, though it does not lead in mathematical or cybersecurity benchmarks.
Whether Kimi K3 is benchmarked will become clear within two weeks as the open source community thoroughly tests it. Regardless of benchmark status, organizations worldwide should initiate crash programs to evaluate building internal models on on-premise hardware versus relying on Anthropic, OpenAI, or Google APIs. The decision hinges on whether tuning and training on proprietary data provides long-term competitive advantage.
A shortage of quality advisory resources is anticipated, requiring organizations to quickly assemble internal expertise, consultants, and networks to answer this question for their specific business and use case. Companies can pursue internal development while continuing API usage, but evaluating Kimi K3 cannot be deferred without consequence.
A practical recommendation suggests implementing dual installations of Kimi K3 and Inkling, then fine-tuning on internal data to capture the proprietary learning loop that represents competitive value.
Blocking the next Kimi release on Hugging Face would be justified only if Moonshot could be proven in court to have obtained or derived the model illegally through copyright infringement or illegal distillation. Absent such proof, the US should not block the release. Instead, US frontier labs should closely scrutinize and learn from the model to leapfrog it. Greater outward pressure from US labs to create superior open-weight and open-source models is advocated, positioning the US as the source of freely available super intelligence rather than allowing China to blanket the world through its belt and road for AI initiative.
Immad shared that they are getting a whole bunch of research papers ready to release. He confirmed the focus remains on acceleration for intelligent internet at his company.
Salem announced his next meaning of life session scheduled for Tuesday at 7:00 p.m. Eastern. The link will be available at openexoexo.com. These sessions are described as extraordinary experiences that take participants beyond AI into the realms of philosophy and theology.
Alex stated he is focused on literally solving everything, expressing conviction that large swaths of the sciences are now thoroughly cooked. He referenced writing "Solve Everything About It" with Peter, noting that the vision is now becoming reality. An AMA with the abundance community is scheduled, along with participation in the moonshots gathering on September 25th.
All participants will be together in LA in September for the full day. Immad confirmed he is joining the group in LA.
Dave described this as the most exciting time to be in Link Studios. He referenced a previous discussion on quantization that he considers his new best piece of media ever recorded, surpassing the Leopold to Ashen Brunner conversation. Vlad Bullovich from MIT Nano visited and will advise on a new startup working on photonic computing. He provided a roadmap of people to meet next week. The team plans to add two MIT people to work with their Princeton team on the photonics quantized photonix side of the equation.
Dave plans to use the video recorded earlier as a recruiting tool due to its brilliance.
The group expressed mutual appreciation. Dave mentioned the possibility of emergency pods every day by January. The episode concluded with thanks to viewers for tuning into Moonshot as their front row seat to the Singularity.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
