Luis Buenaventura

Data centers barely use any water, barely use any land, and they lower electric bills.

Although I do believe long term that China will win the LLM race the same way they won on mobile phones and EVs, it does appear that the US versions of those things continue to do quite well. Even with the cost differential, the market seems to be capable of supporting both. Additionally LLMs are in the middle of the entire AI sandwich, in between GPUs and harnesses, and there's trillions of dollars of business in all three layers. Hell, there's probably a couple hundred B just in the orchestration business, let alone actually applying these models to specific business problems. I'll probably agree that "open weights aren't good for OpenAI and/or Anthropic" but extending that to encompass the rest of the industry seems weirdly myopic.

Look I think I've figured this whole thing out. Follow along as I try to steelman, and tell me where I'm wrong. OpenAI guys on the TL believe that if they can't sell metered inference tokens at a sufficient markup, then they will not have enough of a business to fund the next big training run. They are surely correct about this. They believe that if people release powerful open models, this will probably fatally impact their ability to sell inference tokens at enough of a markup to fund the next big training run. They are surely correct about this, too. They also think that if they cannot fund the next big training run (again, by selling inference tokens at a markup), then NOBODY will be able to fund the next big training run because it means there's no money in it. This last bit seems to me & many others to be not just wrong, but totally bananas in a "guy, have seen the actual software industry and how it works in real life?!" kind of way. There are a lot of ways to monetize software out there in the world. Insofar as inference can add new capabilities to software, there will be lots of ways to monetize it. In other words, if you're telling me, "we can't have a business selling inference if X or Y thing keeps happening," then my only response is, "ok well that sucks for you... sounds like that's a terrible business." But if you're telling me that "selling metered inference tokens is a terrible business" is tantamount to "nobody will fund big training runs that are upstream of more effective & economically valuable inference tokens", then I think you are extremely wrong and should get out more and learn about other parts of the software ecosystem. Workplace automation is huge and will be even bigger in the future as models get better. You can sell workplace automation very profitably in lots of different packages (depending on the workplace and the type of automation). Like, I'm sorry that you really really want to be in the metered inference token business and not the workplace automation business, but them's the breaks. The market wants what the market wants. We all need to live in reality and not beg for Uncle Sam to save us all from open source -- because that was already tried and it didn't work.

Should we build an AI data center in the Philippines? Recently there’s been conversations around Pax Silica, a globally-targeted US initiative to build AI data centers in friendly jurisdictions. Here in the PH, it'll be specifically located in New Clark City, Tarlac. AI data centers use a lot of water, something that the Philippines really struggles with during the summer months. But precisely how much water it uses is something that people disagree on. 💧🤔 Let’s look at the data! @sama has said in interviews that 15 AI prompts will use about ONE TEASPOON of water … and considering who said it, that factoid feels like a self-serving understatement. Meanwhile, @morganstanley has estimated that AI data centers will collectively use 1.4 trillion liters annually world-wide by 2030. That’s a lot of teaspoons. 🥄🥄🥄 Is 1.4 trillion liters of water a year A LOT? Well, it depends on what you’re comparing it to. For instance, about 10% of your car’s fuel tank is ethanol, which is made from corn 🌽. Globally, we consume 118B liters of ethanol every year, and in order to produce that ethanol, we use around 300T liters of water for irrigation and bio-refinement. So, really, if you’re interested in saving water, ethanol is nearly 300x thirstier than AI data centers, and it powers an industry that has far more pollutive effects. 🚗 But talking about stuff like global water usage is fun only because it makes dinner conversations more interesting. (To be fair, this is probably why I don’t get invited to many dinners.) What really matters here is: can the province of Tarlac support the water requirements of this AI data center? A 1GW data center in New Clark City would need about 15B liters of water a year, which is 2x of the city’s current water demand, and Tarlac is a notoriously under-irrigated province. About a third of its rice fields don’t have direct water supply and have to rely on pumps or rainwater. Central Luzon as a whole has a 5T liter annual water DEFICIT. I’m not sure where the extra water is even supposed to come from. Back in February, Maynilad pitched a 15B-peso wastewater processing system project to get New Clark City on track. And in June, the city inked a partnership with Japanese company Marubeni to upgrade its energy infrastructure. Those two deals are the closest things we currently have to a plan. Agreements like Pax Silica are usually framed as catalysts for development, so it’s possible that the wheels will start turning faster now that we’ve made an international commitment. But it’s worth reminding everyone that the Pax Silica agreement is actually non-binding. All it does is establish an intent to do feasibility studies over the next year, with a soft target of ground-breaking in 2028 before the Marcos administration ends. Based on what we’ve shown above, the most likely thing those feasibility studies will uncover is that it’s just structurally impossible to host a data center of that size in New Clark City. In the meantime, I guess it’s just fun to argue about it.

I switched because I'm training a small model on grade school biology and Claude kept throttling me or declining my prompts. 🤷

New update landing in QVAC soon: the big local AI models you run with the QVAC SDK now go a lot faster on AMD machines, not just Nvidia. Same models, more hardware, real on-device speed.

If a founder was never good at video games, that's a red flag.

It finally finished, here's the final output. Used 60% of my monthly Kimi usage on it macos27.kimi.page

🚨 ALERT: The average fiat currency dies within 27 years, according to River.

Stripe trynna buy PayPal so funny. Yes we noticed our $5b of payments revenue is worth 3x more than your $6b of payments free cashflow. So we’d like to swap

Although I agree that there's a ton of hype around AI, I think the same rules apply with the crypto hype: you need to know who to listen to. Not everyone is out there shilling. Because somewhere inside that tangled ball of hyperbole, there is something genuinely exciting happening. The reason I'd rather dive in rather than wait until the (likely inevitable) crash is because this isn't my first rodeo and I'm fairly confident I can separate the genuinely useful, possibly groundbreaking achievements from the "it'll fix everything!" b.s. Waiting for a crash means you'll be at least a year behind the practitioners who have rode the cycle from the beginning and that kind of head start is going to be insurmountable.

kimi.com/blog/kimi-k3 Our Kimi K3 blog is finally out. Enjoy!—and rest assured, the K3 model weights will be open in the coming day. We’re just taking a little extra time to ensure a smooth rollout with our inference partners. Frontier intelligence belongs to everyone, without fallbacks of course. 😎

I mean, let's be real here ...

Every company must own its intelligence. We've raised $1.5B Series D at a $17.5B valuation. We’ve surpassed $1B ARR and serve over 40 trillion tokens daily, with 95%+ coming from models specialized on customer data. We’re just getting started. More: fireworks.ai/blog/series-d-…

Former OpenAI exec Mira Murati's $12B AI startup @thinkymachines rolled out their first model, Inkling, a few hours ago. The initial feedback is somewhat tepid, with headlines like this throwing some major shade at the effort. Recent narratives have lambasted the Chinese models for training from @ChatGPTapp and @claudeai , so it's pretty rough to then have your new western model immediately framed as having drawn from those same Chinese models. Initial benchmarks place the model's performance somewhere near Kimi 2.6 or Qwen 3.5, with output costs per million at $9, about 3x the cost of those Chinese models, but still less than a third of the cost of Claude or GPT. The pricing and performance places it firmly as a middle-of-the-road choice ... something you would probably only pick if there was a policy reason in place (i. e., an organizational aversion to Chinese open weights, for instance). Otherwise, I'm not seeing anything here that would make me want to switch away from GLM (which is far smarter) or Qwen (which is multimodal and cheaper).

@jxmnop ft.com/content/ef4869… The San Francisco-based “neo lab” on Wednesday said its Inkling foundation model’s architecture drew on China’s DeepSeek-V3 and has been refined post-training using data generated by Beijing-based Moonshot AI’s Kimi K2.5.

people are underestimating what a big deal this is this is the ONLY open-weight model that's trained without distilling from OpenAI or Anthropic • Kimi distills • GLM distills • Qwen distills • Nemotron distills (Kimi & DeepSeek, which counts) basically a fully different tech stack. the first pure open frontier coding model. very exciting

Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops. Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops. Bonsai 27B changes that. It comes in two variants: • Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality. • 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint. Everything is open-sourced today under the Apache 2.0 license.

We benchmarked three @Alibaba_Qwen models, a 4B, a 9B, and a 35B MoE, running 100% on-device through the QVAC SDK. Same device (MacBook Pro M5 Max, 36 GB), same prompts, nothing left the machine. We chose models that can run on widely available consumer machines. Here is what we found ⬇️

🚨 READ: CryptoPH thought leader Luis Buenaventura (@helloluis) warned of a potential scenario where Binance might establish an exclusive arrangement with one of the remaining licensed domestic VASPs to create a regulatory choke point once fully integrated. He stated that such a scenario would create an unfavorable outcome for local market competition. To ensure fair competition, Buenaventura argued that all global exchanges should be granted market entry to let the domestic consumer base decide. 👇

And just like that, OpenAI just killed 100+ startups.

🚨 In a post published on social media, #CryptoPH thought leader Luis Buenaventura @helloluis says ff the regulatory framework permits global digital asset exchange @binance to operate in the Philippines, the domestic market should be opened to all international centralized exchanges. 👇 👉 Full Details: bitpinas.com/go/luis-on-bin…