ComputeMaxi (@ComputeMaxi) header
ComputeMaxi (@ComputeMaxi) profile picture

AI enthusiast. Interested in early stage AI/compute infra. Not financial advice.

United States
Joined September 2022
Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)Photo by ComputeMaxi (@ComputeMaxi)
ComputeMaxi retweeted
Awesome open-weight release by @thinkymachines! It's ranked a tad above MiniMax M3 at half the cost per task.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index. thinkingmachines.ai/news/ink…
6
3
44
1,971
ComputeMaxi retweeted
Ironically AI stocks getting killed could push people to free local AI inference because the labs longer have the funds to subsidize AI usage at scale
14
3
1
21
2,885
ComputeMaxi retweeted
AWS Marketplace just added x402. Agents pay per request via Bedrock AgentCore, and it lands on your AWS bill in USD, removing the need for wallets. x402 everything
19
10
1
119
4,079
ComputeMaxi retweeted
soon
8
3
1
65
8,599
LLMs don't predict the next word. They predict the next token. This means they can understand basically any structure that can be tokenized. Not just language, but pictures, video, sound, smell, gravitational orientation, force feedback, etc. it literally doesn't matter.
Geoffrey Hinton: The argument that LLMs are just predicting the next word falls apart when you test their ability to understand something they've never seen before. “I have a former graduate student who's very smart and a very good programmer. And he did a little experiment.” “He defined a new programming language with very unusual syntax.” “He then gave it a program and said, what would this do? And it answered correctly.” “It could understand the definition of a new programming language and figure out what programs in that language would do.” “The idea that it's just predicting the next word doesn't make any sense in that context.” “It had to understand what was going on.”
61
117
12
1,972
85,076
ComputeMaxi retweeted
It's pretty funny right now has AI has turned into a financing story, and Silicon Valley doesn't understand finance, and the stock market has become an AI story, and Wall Street doesn't really understand tech.
57
166
35
1,931
157,315
ComputeMaxi retweeted
an insane release by Google
NEW: Gemini 4 Neon is the highest-performing model we've measured. Not only is it the most intelligent model on our leaderboard, it's also up to 10x cheaper than the next best, reshaping the Pareto frontier of cost vs. intelligence on web research tasks.
6
7
414
15,027
ComputeMaxi retweeted
Request for startups: I am looking to invest in frontier entrepreneurs building in any of these areas. 1/ AI mediation: One AI system does the work for both sides of a legal case and comes up with the settlement. 2/ Polymarket for insurance: You should be able to bid on offering insurance to people. For example, many homeowners in the Pacific Palisades can't get fire insurance anymore. Build a platform where people bid to offer that insurance, and let the bids price the risk. 3/ Interspecies Communication: We're looking for AI that helps us talk to animals. Decode what they are saying, and let us talk back. 4/ New countries: Anyone starting a new country with a new model for how it runs. 5/ Electromagnetic rocket launches: Build a system that launches rockets horizontally, then vertically, using electromagnetic force instead of traditional propulsion. These ideas are just early concepts. We're open to anything along these lines, so I'd like you to pitch us why your angle is the right one. Comment below, and send me your pitch at [email protected]
308
142
57
1,398
114,105
KYC for GPUs is gonna do more damage to American AI than anybody else, and certainly won’t stop John Whiteguy from Obvious Chinese Shell Company Inc from getting access to clusters
14
27
3
386
9,742
ComputeMaxi retweeted
"Compute is the new economy, and tokens the new energy unit."
I’m being pitched a new inference neocloud company on average every two days. No exaggeration. I take every entrepreneur seriously and I look deeply into each. But it’s hard to fathom this being sustainable. At the same time, AI Gateway token demand on Vercel has been ~40% MoM compounded since the beginning of the year. No signs of stopping. One way to reconcile this is that compute is the new economy, and tokens the new energy unit. There’s a new Industrial Revolution taking place. Every aspect of our lives is being replatformed on top of this new system that enhances everything. More like an “oil rush” than a “gold rush”: it’s durable, not a fad. In the short term, I anticipate more expansion, followed by consolidation due to economies of scale, cybersecurity and enterprise requirements (ZDR and cyber being a hot topic for inference providers), capital requirements, and an evolving regulatory landscape.
3
50
5,400
ComputeMaxi retweeted
What is happening??? Google releasing models like CRAZY
tomorrow DeepMind will be dropping Gemini 4 Sodium, a fast yet powerful frontier model which costs just $11.75/M (introductory pricing, raises to $16.50 soon) and is their best model for various things and stuff, rivaling Argon, Carbon, Hydrogen, and Neon in several situations
85
22
6
1,294
75,190
ComputeMaxi retweeted
Samsung killed the GPU mafia. They open-sourced a method that shrinks 70B LLM into less than 1 bit per weight Running a frontier model meant paying API tolls or renting racks of $30,000 GPUs in the cloud. First, we shrank models from 16-bit down to 4-bit. Then came 1-bit models. Everyone assumed 1 bit was the hard floor. After all, how do you go lower than a single 1 or 0? But, researchers introduced NanoQuant, the first ever sub-1-bit LLM compression. Instead of storing every weight individually, even as a single bit, it compresses the model's weight matrices into ultra-compact, structured binary representations. Less than one bit per parameter. On average. The implications are massive: • Slashes VRAM requirements by over an order of magnitude. • Breaks the memory bandwidth bottleneck that makes local inference slow. • Brings massive, high-capability models down to consumer hardware and edge devices. Big Tech built its entire business model on one assumption: that frontier AI is too heavy for you to run yourself. That assumption is dying.
43
106
9
684
25,570
ComputeMaxi retweeted
Something is going on at Google I think they might have hit RSI They’ve released new frontier models FOUR TIMES in less than a month. And each is an actual MAJOR improvement Carbon is the most intelligent model I’ve ever used While everyone is talking about Anthropic and OpenAI, Google quickly caught up What I absolutely love about these models too is they were clearly trained in a completely different way than Claude or ChatGPT. 0 distillation. They talk, think, write, and code COMPLETELY differently than any other model I can always count on Carbon to give me a novel insight no other model has given me If you haven’t incorporated Gemini into your tool belt yet, you absolutely need to
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
211
243
34
3,897
319,116
Supercloud 🤝 Supercycle 🤝 Superintelligence
A lot of exciting AI stuff brewing this week > Luxor, Architect & Hyperbolic did the first $5M OTC put option on B200 rental prices (Hyperbolic covering its downsides in the case that B200 rental price falls) > Akash announced Akash Compute Forwards, letting buyers and sellers lock in the price of future compute (GPU hours) today, with physical delivery of attested hardware later. Pilot scheduled to open in March 2027 > Pearl got listed on OpenRouter, already serving over 20B DeepSeek 4.1 flash 1017 on the first day > Compute Desk raises $45M from LightSpeed to build hedging infra for compute > Mega (prev. MegaETH) also has inference futures as part of their roadmap Everybody is building the financial infrastructure to support compute as the next frontier asset Exciting times
1
6
94
2,651
You have absolutely no excuse not to download AI while you can on your own computer. Don’t have one? Get some junk no one wants and follow my steps. You may have just hours left…
Local AI On An Old 2017 Computer I show you how to run a private AI with no GPU on a PC or Mac from 2017 or later: a beginner's install guide, a computer-age chart, model picks, and the Raspberry Pi verdict. Do it while you still are allowed to: readmultiplex.com/2026/11/18…
27
51
2
392
20,429
ComputeMaxi retweeted
Literally the most important skill for anyone with interest in local AI to use. If you work on inference this is my crown jewel. Open source must win.
I've made 100,000$ with this skill, at least. How can I keep these things to my self? Open source must win. github.com/sybil-solutions/i…
15
24
3
587
24,306
ComputeMaxi retweeted
Open weights are open. Anyone can download them, test them, fine-tune them, and run them on their own hardware. They’re the most inspectable kind of AI there is. A model’s country of origin doesn’t tell you whether it’s safe or good. Benchmarks and evals do. and open source has never had a passport. Linux started in Finland, Python in the Netherlands, Ruby in Japan, Lua in Brazil, llama.cpp in Bulgaria. Kernel commits come from Intel, Google, Red Hat, Huawei and 2,000+ developers each release. Nobody puts a banner on their site about that. Same with open models and the tools that serve them: Mistral from France, Falcon from the UAE, Qwen and DeepSeek from China, Llama and gpt-oss from the US. Roughly 33% of vLLM’s contributors are based in China, and about 52% for SGLang, so need disclosure for all of these also? Huawei was the top employer by kernel changesets in Linux 5.10 and 5.14. I can go on.... No one puts a banner on their site about those. Most technology stack is already "multinational" with open contributions for public use.
SITUATION DETECTED: The White House and Treasury are discussing a proposal that would require US companies to publicly disclose the use of Chinese open-weight AI models on their websites, aiming to discourage their use through public pressure, per Reuters.
11
64
7
407
19,919
ComputeMaxi retweeted
And this is basically what we do every day with people
A slide from @AnthropicAI's IPO roadshow. Hits hard and is very true. AI is soooo much better than how most people are using it.
4
2
29
2,443