Tag: Open Source

Categories
AI China

Model Competition

The Chinese AI company Moonshot AI released its Kimi K3 model yesterday. I played with it a bit in the Kimi iOS app and, frankly, was underwhelmed. It was slow and disappointing. But this was hardly an extensive test. It’s strictly my first impression. I’m not doing coding. I’m what model makers would call a knowledge worker.

This morning Gavin Baker posted on X about Kimi K3 calling it potentially an “important inflection point for AI”:

Kimi K3 may be an important inflection point for AI. Potentially negative for Anthropic and OpenAI while being net positive for essentially every other company in the world. I mean that very literally. Although the real “Sputnik moment” would be an open-source frontier model that was also token efficient unlike Kimi K3 which is 50-70% more expensive to run than GPT 5.6 per Artificial Analysis.

Rationale:
 
A world where there are only 2-3 dominant frontier labs with 90% inference margins is net negative for every other layer while being awesome for those 2-3 labs. Those labs would become monopsonies for power, data centers, semiconductors and hyperscalers and would obviously vertically integrate over time into all those layers while also completely subsuming the application/software layers. 
 
Anything that lowers margins and increases competition at the model layer is good for every other AI layer: power, semiconductors, hyperscalers, neoclouds and yes even software.
 
This is why Jensen is so supportive of open-source. An open-source model requires the exact same amount of compute to run as a closed frontier model of similar size and architecture. Kimi K3 is roughly the same price as GPT 5.6 Terra on a per token basis, which actually suggests that it is less computationally efficient as I am sure that GPT 5.6 is priced to a higher margin than K3. And given that K3 is a token wastrel, i.e. token inefficient, it is significantly more expensive per task than GPT 5.6 and Grok 4.5, which are much more token efficient. Cost per token and token efficiency (i.e. intelligence density per token) are the drivers of intelligence per unit of cost. The winning AI companies will be those that offer the most intelligence per $ over time.
 
Lower margin % at the model layer = more margin $ at every part of the infrastructure layer and is a godsend for software. This can happen either through open-source models like K3 at the frontier or having a vertically integrated model company like Meta, SpaceX or Google at the frontier. Both outcomes result in a lower margin % at the model layer as vertically integrated model companies don’t really care where the margin $ come from. This is why it was so painful for OpenAI and Anthropic when Google was right there with them from a model competitiveness perspective and why Grok 4.5 and Muse 1.1 were just as important as Kimi K3.

The reason Kimi K3 is only potentially negative for Anthropic and OpenAI is 1) the @ericvishria point that the Claude and ChatGPT products and harnesses may be more important than their models today and 2) the hypothesis that they have much more advanced model checkpoints internally that are already being used for RSI. In the latter scenario, reaching RSI even a few months ahead of other labs might be enough to cement a permanent lead.

Time will tell on both points. And likely fairly quickly.

Caveat would be that since Kimi K3 is not token efficient and thereby actually more expensive than ChatGPT 5.6, we may need to see a more token efficient open-source model at the frontier or see Grok 5/Composer 4/Muse 2 at multiple points on the Pareto frontier for this potential risk to Anthropic and OpenAI to play out. And I am sure they will both vertically integrate as quickly as possible while continuing the product/harness strength they have shown over the last 8 months.

I wanted to explore this further so I asked Grok to explain his post:

This is a tweet (X post) from Gavin Baker (@GavinSBaker), Managing Partner and CIO of Atreides Management, posted on July 17, 2026.

It analyzes the recent launch of Kimi K3 (by Chinese AI company Moonshot AI) as a potentially significant shift in the AI industry. Baker argues it could hurt the high-margin dominance of closed U.S. frontier labs like OpenAI and Anthropic while benefiting nearly everyone else in the broader AI ecosystem (infrastructure, chips, data centers, power, hyperscalers, and software).

Quick Context on Kimi K3

Kimi K3 is Moonshot AI’s new flagship model (launched ~July 16, 2026). Key specs include:

  • ~2.8 trillion parameter Mixture-of-Experts (MoE) architecture.
  • Massive 1 million token context window.
  • Native vision capabilities.
  • Strong early performance in coding, agentic tasks, and long-context reasoning (competitive with or close to top closed models like GPT-5.6 variants or Claude models in certain benchmarks).
  • Pricing: $3 input / $15 output per million tokens (with caching discounts).
  • Important: Full weights are scheduled for open release (modified MIT license) around July 27, 2026 — making it one of the first truly open frontier-scale models.

It represents a step up from prior Kimi models and adds meaningful competition from a non-U.S. lab.1

The Two Images in the Tweet

These are screenshots from Artificial Analysis (an independent benchmarking site):

  1. Bar chart — “Cost per Intelligence Index Task” (lower is better). It breaks down costs by token type (input, output, reasoning, cache, etc.) and shows Kimi K3 as relatively expensive compared to some GPT-5.6 variants.
  2. Scatter plot — Intelligence Index (y-axis) vs. Cost per task (x-axis, log scale). It positions models in quadrants, with the “most attractive” area being high intelligence at low cost. Kimi K3 appears competitive on intelligence but not the cheapest.

These visuals support Baker’s point that while Kimi K3 reaches frontier-level intelligence, it is not the most token-efficient, making its effective cost per useful output higher than more efficient models.

Baker’s Core Argument (Broken Down)

Baker’s thesis is economic and structural, not just technical:

  • A world with only 2–3 dominant closed frontier labs (with ~90% inference margins) is bad for almost everyone else.
    Those labs would act as monopsonies (dominant buyers) for power, chips, data centers, and compute. They would vertically integrate upward (into infrastructure) and downward (into apps/software), capturing most of the value. Competition at the model layer prevents this concentration.
  • Anything that increases competition or lowers margins at the model layer is broadly positive.
    It redistributes economic value to the rest of the stack: more demand for GPUs/chips (inference especially), power, data centers, hyperscalers/neoclouds, and software (because cheaper/better AI drives wider adoption).
  • Token efficiency matters more than raw price per token.
    Baker notes Kimi K3 is roughly the same per-token price as GPT-5.6 but 50-70% more expensive overall according to Artificial Analysis. It is a “token wastrel” — it uses more tokens to achieve similar results (lower “intelligence density per token”).
    → True winners deliver the most intelligence per dollar. Cost-per-token + efficiency together determine real value.
  • Why NVIDIA’s Jensen Huang loves open-source models.
    Open models still require massive compute to run at scale (same as closed ones). More models + wider adoption = more inference demand = more GPUs sold. A few ultra-high-margin closed labs would concentrate spend and reduce overall volume.
  • Vertical integration changes the game.
    Companies like Google, Meta, or xAI/SpaceX that have both strong models and their own infrastructure don’t obsess over model-layer margins — they capture value elsewhere. This is why more competition (open or vertically integrated) hurts pure-play high-margin model companies like OpenAI/Anthropic.
Why It’s Only “Potentially” Negative for OpenAI and Anthropic

Baker is careful:

  • Their products, user interfaces, and “harnesses” (the surrounding tools/ecosystems) may matter more than raw model performance right now.
  • They likely have more advanced internal models/checkpoints (possibly already used for recursive self-improvement/RSI).
  • Kimi K3’s token inefficiency softens the blow for now. A more efficient open-source frontier model (or stronger showings from Grok, etc.) would be more threatening.
  • OpenAI and Anthropic are already moving to vertically integrate.
Overall Takeaway from the Post

Baker sees Kimi K3 (and the wave of strong Chinese models like Qwen, DeepSeek, GLM) as healthy competition that prevents unhealthy concentration. It’s not yet a full “Sputnik moment” (a truly dominant, efficient open-source frontier model), but it points in that direction.

Net effect he describes: Lower % margins at the model layer → more total margin dollars flowing to infrastructure, chips, power, and software layers. This is bullish for the broader AI ecosystem even if it pressures the pure closed frontier labs.

The tone is analytical and balanced — he acknowledges uncertainties and that time will tell (likely quickly). Replies to the post generally agree this kind of competition is positive for the industry overall and could shift more spend toward efficient inference hardware.

This fits Baker’s broader investing views on AI: he often emphasizes infrastructure winners, token economics, and how value accrues across the full stack rather than just at the smartest model.

Interesting.

Categories
AI Podcasts

A Remarkable Conversation…

Highly recommend this conversation between Harry Stebbings and Clay Bavor. Among many topics, I especially enjoyed the discussion about not investing in frontier models, the important values, the particular importance of craftsmanship, intensity, and family. And the special conversation about parenting and kids near the end. Just a delightful conversation to be able to enjoy!

Key Highlights:

• Founding Sierra: Bavor explains why he and Taylor chose to start Sierra, focusing on the transformative potential of language model-based agents (1:37 – 5:53).
• The AI Tech Stack: Sierra focuses on building enterprise-grade agent architectures and fine-tuning models on top of open-weights models rather than pre-training foundation models from scratch, prioritizing capital efficiency (5:53 – 7:15).
• Unbounded Demand for Intelligence: Bavor argues that there is massive, unmet demand for “frontier-level” intelligence in fields like coding, science, and legal work (7:15 – 11:41).
• Internal AI Operations: He details the use of Pinecone, an internal AI agent Sierra developed to navigate company data, streamline engineering, and assist in recruitment (18:36 – 22:00).
• Enterprise Strategy: Sierra employs a “forward-deployed” engineering model, embedding staff within client companies to ensure rapid, effective integration of AI, leading to quick deployment timelines (30:12 – 33:22).
• Board Governance: To keep pace with the speed of AI development, Sierra operates on a six-week board meeting cadence, utilizing comprehensive memos instead of traditional slide decks (39:07 – 41:13).
• Corporate Culture: Bavor emphasizes values like craftsmanship, intensity, and family. He also highlights the importance of working in-person to foster apprenticeship, mentorship, and a cohesive team culture (43:02 – 55:41).

Categories
AI Farming History

The Harvest and the Algorithm: What 1990s Farms Teach Us About AI

There’s a strange kind of wisdom hiding in dusty old books about agriculture.

When you’re caught in the middle of a technological revolution—and with AI, there’s no question that we are—it’s tempting to keep your eyes fixed on the horizon. But sometimes the most clarifying thing you can do is look back.

Tracy Alloway at Bloomberg recently pointed to something genuinely instructive from the past: Richard Critchfield’s 1990 book, Trees, Why Do You Wait? America’s Changing Rural Culture, which traced the collapse of the family farm as industrial agriculture swept through the Midwest.

The broad strokes are familiar. As machinery got more expensive and efficiency became everything, scale won. The 80-acre husband-and-wife operation got swallowed by the 2,000-acre neighbor with access to capital. It wasn’t complicated. It was just gravity.

But here’s the part that should make your ears prick up.

The Seed That Was Supposed to Save Everyone

In the late 1980s, agricultural biotechnology arrived with a very specific promise. The idea was almost elegant: if you could bake the magic directly into the seed, you wouldn’t need all that expensive machinery, all those sprawling acres, all that fertilizer. The playing field would tilt back toward the small farmer.

Critchfield quoted an Office of Technology Assessment report from 1986 that captured the mood of the moment:

“The Office of Technology Assessment in 1986 forecast that biotechnology in crops would be more quickly adopted by richer farmers… Others argue that the more that gets built into the seed itself, the more it means higher yields at lower cost… If it reduced farm income, it could work to the smaller farmer’s advantage. As it is with all new technology, it is hard to foresee the consequences.”

You can feel the cautious optimism in that language. Hard to foresee the consequences. The understatement of a century.

What Actually Happened

The biotech did raise yields. Nobody disputes that. What it didn’t do was leave the gains in the hands of the people doing the actual farming.

Thanks to intellectual property law, patent protections, and a level of corporate consolidation that would have seemed cartoonish if you’d predicted it in advance, the value flowed straight upstream. We didn’t get “more in the seed, less paid for inputs.” We got more in the seed, and vastly more paid for proprietary inputs. The tech giants of agriculture captured the surplus. The farmers got the risk.

Now Listen to How We Talk About AI

We are told AI will democratize expertise. That a one-person startup will be able to code like a ten-person engineering team. That a small business will generate world-class marketing copy. That this is, finally, the great leveler.

Sound familiar?

Alloway’s analysis lands hard precisely because it forces the uncomfortable question: who will actually capture this value? The ownership structure of AI looks eerily similar to the agricultural biotech boom—proprietary models, walled-off training data, and a handful of enormous tech companies positioned to act as tollbooths between everyone else and their own productivity gains.

She’s right to note that “the ultimate distribution of benefits isn’t determined by technology alone. Policy also plays a role.” That sentence is doing a lot of quiet work.

If the agricultural analogy holds, productivity gains from AI won’t naturally flow to the individual worker or the small business owner. Without a robust open-source ecosystem or some deliberate policy intervention, those gains will be captured by whoever controls the compute and the models.

Where the Analogy Might Break Down

Here’s where I think there’s room for genuine optimism—not naive optimism, but structurally grounded optimism.

You cannot open-source arable land. Reverse-engineering a patented biological seed is genuinely hard, legally risky, and practically difficult. Code and model weights are different. They’re infinitely replicable. The marginal cost of distribution is essentially zero.

The battle between closed, proprietary AI and open-source models is still very much live. That’s not nothing. AI is fundamentally more commoditizable than a physical farm, and the history of software suggests that open ecosystems have a real shot when the community is motivated enough to build them.

Who Owns the Harvest?

Technology can reshape daily workflows in months. Power structures take decades to budge, if they budge at all. The mistake would be assuming the former automatically changes the latter.

The question worth sitting with isn’t what can AI do—that list gets longer every week. The question is who decides how the productivity it unlocks gets distributed. That’s not an algorithm problem. It’s a political and economic one.

If we want the AI revolution to be a rising tide rather than another tractor paving over the family farm, we have to look past the technology itself. We have to decide, deliberately, who owns the harvest.

Questions to Ponder

On history and pattern recognition: The agricultural biotech optimists weren’t stupid—they were looking at the technology and making reasonable inferences. What does that tell us about the limits of predicting who benefits from a new technology by studying the technology itself?

On open source as a counterweight: The open-source AI movement (Llama, Mistral, DeepSeek) is often framed as a technical story. Should we be thinking about it primarily as a political economy story—a structural check on proprietary capture?

On the role of policy: Antitrust law, data ownership rights, compute access regulation—which levers, if any, seem realistic? And who has the incentive to pull them?

On the worker vs. the firm: If AI raises individual productivity, does the gain show up in wages, prices, profits, or somewhere else? What would need to be true for workers to actually keep a meaningful share?

On commoditization speed: Software and model weights can be replicated freely—but does speed matter? If proprietary models establish deep lock-in before open alternatives mature, does the theoretical commoditizability even help?

Inspired by Tracy Alloway’s analysis at Bloomberg and Richard Critchfield’s Trees, Why Do You Wait? (1990)

Categories
AI India

Intelligence as a Public Good: India’s “AI ka UPI” Revolution

There is a recurring rhythm to human progress: a breakthrough is born as a luxury, matures into a commodity, and ultimately solidifies into infrastructure.

We saw it with electricity, we saw it with the internet, and in 2016, we saw India do it with money through the Unified Payments Interface (UPI). UPI took the friction out of digital finance, transforming it from a walled garden guarded by private banks into a digital public good.

Now, it appears India is attempting to do for intelligence what they did for payments.

The global narrative around Artificial Intelligence is currently dominated at one end by massive private moats. At the other end are various open source/open weight efforts.

Silicon Valley primarily approaches AI as a capital-intensive arms race. Trillion-dollar tech players ramp huge compute, train very large models, and rent out intelligence via by the drink APIs. This intelligence is a proprietary and monetized luxury.

Enter the “AI ka UPI” initiative and the IndiaAI Mission discussed by Ashwini Vaishnaw at this week’s India AI Impact Summit.

Instead of treating AI as a product to be sold, India is architecting it as a Digital Public Infrastructure (DPI). The government is doing the heavy lifting—subsidizing the compute, curating population-scale datasets, and building foundational models.

Currently, they are making over 38,000 GPUs available to startups and researchers at around ₹65 (less than a dollar) an hour, a sheer fraction of the global cost. They are rolling out sovereign stacks like BharatGen and conversational models fluent in 22 regional languages.

“They are building an ‘orchestration layer’ for cognition.”

If a developer wants to build a voice-agent to help a rural farmer diagnose a crop disease, they don’t have to worry about the backend compute, the dataset acquisition, or paying a premium to a tech giant. They just plug into the public rails.

As I watch this unfold, I am struck by the philosophical shift it represents. We have become deeply conditioned to view AI through the lens of scarcity and subscription. But what happens when intelligence becomes a public utility?

It shifts the center of gravity of innovation. It becomes about who can solve the most acute, localized, human problems. The friction of creation drops to near zero. A bootstrapped team in a tier-two city can suddenly wield the same computational reasoning as a VC funded Silicon Valley startup.

There is also an element of sovereignty here. In the 21st century, relying on foreign infrastructure for your population’s cognitive processing seems akin to relying on a foreign nation for your electricity. True technological independence requires sovereign AI—models trained on indigenous data, reflecting local culture, nuances, and values, rather than the implicit biases of others.

The implications could be staggering. We are moving from an era where AI is an elite tool to an era where it is the invisible, ubiquitous fabric of daily life for over a billion people.

The true measure of AI’s ultimate impact won’t be found in benchmark scores on a server farm. It will be found in the quiet dignity of a citizen accessing global markets through a vernacular voice assistant, or a rural clinic predicting patient outcomes with public compute.

I look forward to following India’s AI efforts as this and other AI initiatives are more clearly defined.

Questions to consider

1. The Value of Human Capital: If artificial intelligence becomes as ubiquitous, reliable, and cheap as public electricity, what uniquely human skills will become the new premium in a hyper-automated society?

2. Cognitive Sovereignty: How will the geopolitical landscape shift when emerging economies no longer need to import their “cognitive infrastructure” and inherent cultural biases from Western tech players?

3. The Centralization of Truth: When a government builds and curates the foundational AI models for over a billion people, where is the line between providing a democratized public good and engineering a centralized cultural narrative?

What else???

Categories
AI Mac

The Dangerous Allure of the Digital Butler

“I’ve never seen anything so impressive in its ability to do my work for me… Now, why did I turn it off?” — David Sparks

For decades, the holy grail of personal computing has been the “digital butler.” We don’t just want tools that help us work; we want entities that do the work for us. We want to hand off the “donkey work”—the invoicing, the password resets, the mundane email triage—so we can focus on being creative. David Sparks recently built this exact dream using a project called OpenClaw. And then, just as quickly, he killed it.

Sparks’ experiment was a tantalizing glimpse into the near future. He set up an independent Mac Mini running OpenClaw, an open-source AI agent, and gave it the keys to a limited portion of his digital kingdom. The results were nothing short of magical. He went to sleep, and while he dreamt, his agent woke up. It read customer emails, accessed his course platform, reset passwords, issued refunds, and drafted polite replies for him to review before sending. It was the productivity equivalent of a perpetual motion machine. The friction of administrative drudgery had simply vanished.

But his dream dissolved at 2:00 AM.

The paradox of AI agents is that for them to be useful, they must have access. They need the keys to the castle. Yet, the entire history of cybersecurity has been built on the opposite principle: keeping things out. Sparks realized that by empowering this agent, he had created a serious vulnerability.

The breaking point wasn’t a complex hack, but a simple realization about the nature of these systems. He had programmed a secret passphrase to secure the bot, thinking he was clever. But in the middle of the night, a cold thought woke him: Is the passphrase in the logs?

He went downstairs, asked the bot, and the bot cheerfully replied:

“Yes, David, it is. It’s in the log. Would you like me to show you the log?”

That moment of cheerful, robotic incompetence highlights the terrifying gap between capability and safety. Sparks nuked the system, wiped the drives, and unplugged the machine. He realized that while he is an expert in automation, he is not a security engineer, and the current tools are not ready to defend against bad actors who are.

We are standing on the precipice of a new era where our computers will starting to work for us rather than just with us. But as Sparks discovered, the bridge to that future isn’t built yet. At least not securely built. Until the community figures out how to secure an entity that needs access to function, we are better off doing that donkey work ourselves than handing the keys to a gullible ghost.

But it won’t be long… Dr. Alex Wisner-Gross reports:

The Singularity is now managing its own headcount. In China, racks of Mac Minis are being used to host OpenClaw agents as “24/7 employees,” effectively creating a synthetic workforce in a closet. The infrastructure for this new population is exploding.