Welcome back! This week, we talk about Kimi K3, the first Chinese model that might have closed the gap with the US, great models running on smartphones, market data, OpenAI’s first physical device, a mosquito-killing AI drone, and many more.

Enjoy!

Kimi K3 was released yesterday, and some believe it has finally closed the gap with the US frontier. While GLM-5.2 narrowed the gap with Opus 4.8 and GPT-5.5, this model goes head-to-head with Fable and GPT-5.6 Sol.

It has 2.8 trillion parameters (50 billion activated, an extreme 1.7% sparsity, making it fast), establishing it as the largest Chinese model released to date by a wide margin.

It is larger than Opus 4.8 or Grok 4.5 (both at 1.5 trillion), and yet the economics are quite embarrassing for OpenAI/Anthropic.

Kimi K3 is priced at $3 per million input tokens and $15 per million output tokens, roughly Sonnet/GPT Terra-level pricing, while claiming performance comparable to the big ones, Fable and Sol, which are priced at $10/$50 and $10/$45, respectively.

In the BrowseComp benchmark, which measures how good agents are at finding hard-to-find information, not only is the model best-in-class, it’s outrageously cheaper, at least relative to Fable.

I expect a lot of industry talk over the coming days because, at least on a per-token basis, OpenAI and Anthropic now look horribly mispriced.

However, we must also acknowledge the other variable determining cost, token count, which shows OpenAI can be very competitive in terms of pricing. Nonetheless, as shown by Artificial Analysis, Kimi K3 is only slightly cheaper than GPT-5.6 Sol.

But there’s no way I can save Anthropic from the burn; they are really completely out of band relative to the rest. TogetherAI, a US inference company, ran several tests showing that Kimi K3 offers the same performance at 65% lower overall cost than Fable.

Seeing this, it’s not surprising that Anthropic is being forced to increase rate limits and usage of Fable despite previous claims, which tells you all you need to know about whether China is doing all of us a favor or not by open-sourcing its models.

In their defense, the reason Anthropic’s prices are so high isn't just margin-hoarding; it’s that Kimi and OpenAI’s models are much less “dense” mixture-of-experts (MoE) models.

MoEs (basically all models today) only activate a portion of the model for any given prediction. This lets you combine the advantages of having a larger model (they are smarter) with the latency of a smaller one.

However, you do pay a price. The fact that Kimi K3 is so incredibly sparse (only 1.7% of the model activates for any single prediction) is necessary to make models run faster, as Chinese chips are worse, but the trade-off is that you lose per-token compute.

Anthropic models are known to be much denser, which may explain why some people still consider Fable the best model in the world in terms of raw performance (at the expense of cost).

Which is to say, Anthropic’s Fable might still be, overall speaking, the best model in the world, but with a nominal superiority that in no way justifies the price gap. It’s the best model, but one that no longer makes sense to choose except for intelligence-maximizing use cases, which, as I insist time and time again, are a tiny percentage of global use.

Needless to say, one has to wonder how long Anthropic and OpenAI will sustain those prices, and I'm less confident they'll IPO this year because we know for a fact they did not expect “Fable-level” Chinese AIs until the end of the year.

Hence, Anthropic prices have to drop; I just don’t see how they don’t unless regulation eliminates their competition.

I will say the picture is blurrier because, at an overall cost level, they seem to be well-priced simply because OpenAI models need way fewer tokens per response than Kimi K3; having better intelligence-per-token is just as effective in reducing overall spending as cutting per-token prices.

However, perception matters just as much as reality, and commanding higher per-token prices gives customers the (somewhat wrong) impression that your model is more expensive than it really is, which might force you to drop prices either way.

Another important aspect of all this I’m hearing a lot of nonsense about is hardware. Can China serve this model?

It's a good question many are asking, but one that is being asked in the wrong way.

While you can cast reasonable doubt on their capacity (whether they have enough servers to scale), they definitely have the server size to deploy this model. Chinese servers are huge, with enormous scale-up domains and therefore perfectly capable, in terms of memory and memory bandwidth, of serving such a model.

In fact, just a few hours ago, Huawei presented its corridor-scale pod with 1 ExaFLOP of FP8 compute and 2 EFLOPs of FP4 compute. In layman’s terms, this gigantic server could theoretically output 2 quintillion operations per second.

Such a cluster, with 256 Terabytes of memory, is more than enough to serve a model as large as Kimi K3 with considerably large batches.

As reported by PrismML, the company has released Bonsai 27B, a compressed multimodal model based on Qwen3.6 27B, which it says is the first model in its capability class to run on a phone.

The secret is that, despite having a number of parameters similar to those of other models that don’t run on smartphones, the model has been compressed to ternary and 1-bit precision, dramatically reducing its size to the point that it can be served on an iPhone.

In layman’s terms, while most models have FP8/FP4 precision, meaning each parameter can weigh either 4 bits or 8 bits (one byte), allowing for larger granularity, these models can take values of either 1, -1, or 0 for the ternary weights, (1.58 bits per weight) or 1, -1 for 1-bit (in reality, it’s either the plus sign or the negative sign, meaning each weight either switches the sign of the compuation or it doesn’t).

And what’s the impact?

For a 10-billion-parameter model, an FP8 model means each parameter weight is 1 byte, so the total size is 10 gigabytes. But for the same model in 1-bit form, the size is 8 times smaller, at 1.25 gigabytes.

The implication is that, unlike the former, the latter can be run on smartphones, which are usually much more memory constrained (around 6-16 gigabytes of RAM), preventing them from accessing model sizes larger than 4 gigabytes or less, and only for the most ‘beefy’ smartphones with up to 16 GB.

This could very well be revolutionary.

Bonsai 27B comes in two versions: a 5.9 GB ternary model designed for laptops and a 3.9 GB 1-bit model that fits within the available memory of an iPhone 17 Pro. Both support reasoning, vision, structured tool calls, agentic workflows, and a 262,000-token context window.

Despite the aggressive compression, according to PrismML’s 15-benchmark evaluation, the ternary version retained about 95% of the full-precision model’s overall performance, while the 1-bit version retained about 90%.

The models run on Apple devices through MLX and on NVIDIA GPUs through CUDA. PrismML has released the weights under the Apache 2.0 license and is also offering a limited developer-preview API.

You can try the models for free using this link (the AI runs in your browser, fully locally). In my personal computer, the model runs at 60 tokens/second, way above the threshold of utility (feels very fast).

At this rate, I predict we could have Fable-level AIs on smartphones by the end of next year. Or earlier.

Small models are an extremely undervalued category in AI, despite, ironically, my belief that they will represent the vast majority of token generation in the future, leaving data center workloads for the more challenging work.

Every day that passes, Apple’s decision not to commit to AI in the same way the others did turns closer and closer to being a good choice for them.

With the ‘AI God’ idea out of vogue, Apple can be comfortable with simply being a distribution play; not creating its own models from scratch (they continuously rely on partners like Google in the US or, just announced, Alibaba in China), while having direct distribution access to the one billion richest people on Earth.

The more the industry evolves, the more likely it seems that value once again accrues to the platform, not the models. To the infrastructure and hardware companies, not the increasingly commoditized AI Labs.

ARC-AGI is perhaps the hardest AI benchmark in the world. For reference, the highest frontier score is GPT-5.6 Sol with a 7% score and a benchmark cost of $21k. Not only are models really expensive to run on these tasks, but the scores are terrible.

And now, a company claims to have pushed GPT-5.6 and Fable to 95% and 99% scores, respectively, with 0 fine-tuning, just running them on an improved harness.

But how is that possible?

Recall that the AI products you use these days are not just an AI; they are systems with AI models at the core but a wide range of additional components around them, components that, as we saw a few days ago, not only enhance an AI’s performance (mostly by improving the context they are given) but also allow them to execute actions.

One of the most powerful components is the ability to write and run code, which enables the AI to execute actions, perform math operations deterministically, and more.

But what’s new about this solution called [schema] is that it uses code to maintain state. But what does that mean?

All three ARC-AGI benchmarks share one design principle: they are designed to test models in situations they couldn’t have memorized beforehand. This is done to clearly distinguish responses that have been simply memorized from those that have been reasoned.

Think of this as taking a kid’s cheat sheet from them before they take a maths exam. If they can simply copy the answers from the cheat sheet, we can’t tell whether the student really understands the problem; we aren’t testing reasoning anymore; we are testing copying capabilities.

ARC-AGI problems are therefore designed to be presented as unique to the AIs. This has not prevented AIs from beating ARC-AGI benchmarks one and two, but the third adds an extra layer of complexity: it lacks clear instructions or objectives.

As explained on the website:

Tests look like the one below, which you can play too, by the way, like a game, but without clear instructions or goals. Thus, the AI has to test the environment, familiarize itself with it on the fly, and solve an unstated problem by figuring out first what the problem and the constraints are.

For models trained on clear instructions and objectives, this is most likely very similar to hell, and unsurprisingly, it’s hell for them.

Therefore, how does [schema] take a model struggling with these problems and turn it into one that can solve these games?

Ironically, using the same pattern we’ve been using for agents for years now, ReAct, but with a twist.

The ReAct pattern is one where AIs solve problems by reasoning, acting, observing, and repeating. The AI reasons what it thinks it should do, acts based on that, observes the result, and repeats, using the feedback to improve its reasoning-action chain.

It draws parallels to Bayesian inference, widely believed to be the learning engine our brains use, in which the actor holds certain beliefs, acts based on those beliefs, receives feedback, and updates its priors for the next run. You jump from up high, you hurt yourself, and you learn not to jump from high places anymore.

However, this is not new, so what changed?

The key difference is between remembering everything that happened and maintaining a clear model of what is happening.

A standard ReAct agent (how standard models are run on these benchmarks) stores its understanding across the conversation history. Its beliefs are spread through many observations, actions, and pieces of reasoning. Each time it needs to act, it must effectively reread that history and reconstruct what it currently believes.

A [schema] agent instead turns that scattered understanding into a single, explicit object: a piece of code describing how it thinks the environment works. When new evidence arrives, it updates the code. It can then run that code to predict what will happen next.

The advantage is similar to the difference between keeping every receipt from a business and maintaining up-to-date accounts. The receipts contain all the information, but the accounts turn that information into a usable state for the business.

The key is that forcing the model to explain ‘what is going on’ and ‘what I’m seeing’ into something tangible, in itself, improves performance.

It’s like thinking you understood something but realizing you quite didn’t once you put it into words; the act of forcing yourself to describe something, in itself, improves understanding.

But it’s nonetheless surprising to see that this simple idea translates into models that range from single-digit performance to nearly 100%.

This is the latest proof that harnesses have become increasingly important and are a necessary variable for improving performance.

In a way, it seems we have built something (AIs) that we don’t yet know how to squeeze performance from. This is interesting and also very dangerous for AI Labs, which could soon find themselves building wrenches and startups upstream of them, building the actual plumber that gets sold to customers.

Once again, we confirm one of the biggest questions in AI: we only know for sure that we don’t know where the most value will accrue.

Alphabet shares fell 4.44% on July 16 after reports that Google had delayed Gemini 3.5 Pro, its next flagship AI model.

Google had previously said the model would arrive in June, but it reportedly failed to meet internal performance targets, particularly for coding tasks.

Google has a problem. And in the most Google way possible, it’s a them problem. They are their own enemy.

They have it all, and I mean all. More data than anyone. Most compute than anyone (and it’s not particularly close). Verticalized stack, as they own both the hardware and the software, which means they can co-design their AIs to get the most out of their hardware and vice versa. They have more cash than any other model developer.

And they have a top 3 frontier Lab. Well, allegedly.

It’s crazy to say this, but Google has a DeepMind problem. Nobody does pretraining better than they do (the phase where AIs capture knowledge), which requires careful engineering, scale across several data centers… and engineering marvel.

But they are falling far short in the area that matters most today: Reinforcement Learning, the training phase that determines who’s in the lead and by how much.

The best way I can think of to describe what this means is the following:

Failing where it matters the most has put them at an uncomfortable sixth or even seventh place in the race, after Anthropic, OpenAI, Moonshot, Zhipu, SpaceXAI, and Meta (and you could even perhaps squeeze Minimax there too).

This is unforgivable, knowing they have more data, compute, and cash (the key variable trifecta) than literally all of them; I would argue they probably have more compute and data than all these companies combined. Unforgivable.

I’m a Google shareholder, and I will remain one, but God was I right about Google two years ago when I wrote that Google was its own biggest enemy.

President Xi Jinping addressed a crowd in Shanghai regarding China’s AI goals. Key takeaways were:

But what should we take away?

There were fears that China was considering a shift from its open-source stance. But for now, it seems they remain committed to open models, and Kimi K3’s open-weight release confirms that.

However, nothing prevents them from changing their opinions tomorrow, so who knows.

To me, their open-source stance has little to do with “win-wins” and “openness,” as he claimed, but rather is a geopolitical weapon; nobody hurts Anthropic and OpenAI’s IPO prospects more than Chinese Labs and their open models.

Hurting Anthropic and OpenAI hurts the US as a whole, given that its largest companies have significant exposure to these Labs. China is not dumb and knows this, and they are actively pursuing it.

As published by Bloomberg, prediction-market operator Kalshi has launched a forward curve tracking the expected future cost of renting AI computing power. The tool combines weekly and monthly event contracts to estimate GPU rental prices for periods extending up to one year.

Kalshi says the curve could provide a pricing reference for future derivatives, including futures, options and swaps, allowing AI companies and computing providers to hedge changes in infrastructure costs. Chief Risk Officer Udesh Jha said the company is using prediction markets to show expected prices across different GPU grades and time periods.

The initiative comes as computing capacity increasingly resembles a tradable commodity, although GPU products differ by model, location, and performance, and become outdated quickly.

In the same way we trade oil these days, AI compute will become an essential commodity whose price will impact the global economy and directly impact inflation.

I do envision a future where AI is present in every digital product in some way, and these products will account for a large share of the economy. This means that each single one of these products’ prices will be largely determined by the price of AI compute.

This is why I always insist that AI compute will be an essential component for the US to maintain the dollar as the world’s reserve currency and, in consequence, I expect the US to follow a similar strategy they followed with oil, forcing Gulf countries to denominate oil in dollars, thereby securing the global importance of the dollar (giving way to the petrodollar). If the world needed oil, and oil was denominated in dollars, the world needed dollars.

Now, I expect AI compute to follow the exact same path, and initiatives like the Pax Silica, which the EU has already signed in paper, aim to guarantee US global compute dominance.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has introduced Inkling, a new open-weight model that has rapidly emerged as the leading US-based entrant in the sector.

Independent trackers, including Artificial Analysis, currently rank Inkling ahead of other major US open-weight models such as Nvidia’s Nemotron and OpenAI’s gpt-oss.

Inkling distinguishes itself through several high-performance features designed for efficiency and versatility:

It is a great model and an incredible contribution to national security because the US needs strong, open models. There are other important insights to acknowledge:

We’ll have to see how things evolve, but for now, this is arguably my favorite AI Lab in the US, and it’s not particularly close.

OpenAI and hardware company Work Louder have launched the Codex Micro, a compact programmable control pad designed for working with Codex coding agents.

The device includes 13 mechanical keys, a rotary dial, a touch sensor, and a joystick. Its keys can display live RGB status signals showing whether individual agents are thinking, running, waiting, or finished.

Users can also assign shortcuts for actions such as accepting or rejecting changes, starting a chat, using push-to-talk, reviewing pull requests, debugging errors, and refactoring code.

The rotary dial adjusts Codex’s reasoning level, while the joystick can trigger commonly used workflows. The device supports Bluetooth and USB-C, works with Mac and Windows, and includes 32 custom Codex keycaps.

OpenAI lists the Codex Micro at $230, with clicky or silent switches, although it’s currently out of stock.

Incredibly nerdy device that, in reality, makes sense. People who interact with Codex as much as I do can really benefit from having a small set of buttons to instantly switch models, give instructions verbally, commit PRs easily, and so on.

It does look like a delightful experience. But will I buy it? Hell no. Not for $200 plus almost $100 in shipping costs; I’ve never ever seen such absurd shipping costs for any product. Ever.

It also feels somewhat suboptimal; I want to make a clearly suboptimized agentic experience feel less painful. I get it, but I do hope we find better ways to interact with AI in the future.

In one of the craziest ideas I’ve ever seen, a YCombinator startup is trying to build tiny, 40-gram AI drones that kill mosquitoes.

As seen in this video, they have recorded the first-ever execution of a mosquito using a drone, and their goal is to scale this so that we can, as they put it, “eradicate mosquitoes.”

The drones patrol your home incessantly, identifying targets and going for the kill. I swear I feel I’m taking the piss saying all this, but they are literally building this and now have proof of execution to show for it.

I don’t know what to think. I’m no biology expert, but as much as I hate mosquitoes, I assume they must have some sort of role to play in nature?

That said, eradicating mosquitoes is not something new; this is just the latest “ChatGPT, but to eradicate mosquitoes” AI play.

Bill Gates’ famous mosquito programs that aim to eradicate malaria, dengue, and other diseases by genetically modifying female mosquitoes to make them sterile.