just some guy

@JigarShahDC @Longviewres Uh are you just making this list up based on the 2022-2024 hype lists + nuclear? There is no universe where Leap, e-zinc, Charm, Electric Era, and Electric Hydrogen successfully raise in public markets in near future. Anth, SpaceX, OAI going to be massive liquidity vacuums

@aakashgupta @gdibner AI slop bots replying to your AI generated tweet. This site is hell

Khosla just paid $1.5B to short the idea that model lock-in is a moat in AI coding. Factory's valuation went from $300M to $1.5B in 7 months. 5x. Look past the number. What Khosla is actually buying is the only company whose core bet is that the foundation model under you stops mattering. Every AI coding platform had to pick a thesis. Cursor: we'll rewrap whichever model wins. Claude Code: our model is the best. Cognition's Devin: we own the agent end to end. Factory's bet is sharper. Agent design beats model choice, and they'll prove it on every frontier model simultaneously. They did. Droid hit #1 on Terminal-Bench with Claude Opus at 58.8%. Then #3 with GPT-5 at 52.5%. Then Sonnet at 50.5%. Three of the top five agents on the hardest end-to-end coding benchmark are all Factory running different models. Claude Code running Claude itself came in at 43.2%. That's the thesis trade. If agent framework beats model selection, then Anthropic and OpenAI get commoditized in code the same way AWS commoditized server hardware. The moat moves from "which model" to "which orchestration layer sits between the developer and the model." Run the math on where the money is going. Cursor is at $29.3B. Replit is at $9B (also Khosla, tripled in 6 months). Cognition, Magic, Codeium, and Factory bring the AI coding stack to roughly $50B in private valuation. The space is being priced like one of them wins a generational prize. Factory is the only one in that set whose product gets better as the model landscape gets noisier. Every new frontier model release is distribution for them. Every model release for a rival is a feature migration risk. The part nobody's pricing: enterprise buyers are starting to ask which vendor survives three years from now. At MongoDB, EY, Bayer, Zapier, and Clari, Factory is already the answer. 31x faster feature delivery and 96.1% shorter migration times is what a CIO shows the board when moving a dev org off one vendor. The real question for the rest of the stack: what happens to your valuation when model choice stops being a purchase criterion?

@TheAhmadOsman I would love your take on the Span Xfra

@RichardMeyerDC I can't get my head around the new Span Xfra thing for this exact reason...

I don't think this is right. Training still wants giant, tightly coupled clusters, and inference increasingly does too. AI still scales with compute. Net of efficiency gains, energy is the constraint. At some point scaling gets uneconomic. That’s a cost limit, not architectural.

@dillon_dunteman @hyperion__cap I would love to read the valar memo. Right now I have them in my promotion and social media bucket. I would love to be wrong though!!

Today, we are unveiling @hyperion__cap, an investment firm built to be the best strategy partner for deeptech founders. Again and again, we heard from these founders that venture capital has been failing them, even as more deeptech funds entered the market in recent years. Too many deeptech VCs lack a real command of industry history, hardware unit economics, go-to-market, and engineering nuance. Instead of rigorously evaluating complex frontier technologies, they often pass with vague references to “science risk." They overlook exceptional founders outside usual elite networks and concentrate capital based on pedigree, reducing their thinking to hand-wavy “founder bets." These VCs then prioritize promotion and social media over delivering real value to founders. In this world, LPs are also losing. And more of them continue to be disappointed by the lack of rigor that their deeptech GPs bring to evaluating these startups. We raised $35 million for Hyperion's Fund I to change this paradigm. On average, we complete 100+ pages of deep research and strategy ideas that are shared with our founders. We also share this industry research with our LP base, which has already helped support our founders with additional capital and valuable introductions. We hold regular strategy sessions with our founders and obtain key connections that unlock new growth vectors for their businesses. Over the last 6 months, we’ve already invested $9 million behind founders across 7 companies: @FarisSbahi at Normal Computing, @isaiah_p_taylor @ Valar Atomics, Will Wilson @AntithesisHQ , @drauwsy @ Kunin, @abeirami & @aparandehgheibi @ [stealth], Charlie Cheng @ TC Lab, and Mike & Josh @ F-ADA. I'm deeply grateful to the senior leadership at Vista Equity Partners for their support, to the venture GPs who have advised us, and to the founders who chose to partner with us in the earliest days. We’re especially grateful to our limited partners, who put their trust in us at the firm's inception. Lastly, to build this firm alongside one of my closest friends and college teammates @henr56520 been a true privilege. We’re looking forward to working relentlessly for the founders we've backed and for those we’ll have the chance to support in the years ahead. hyperioncap.co

@RaoulGMI Hahahah dude. You were actually out in front on the Covid crash. Cuz you were short AF on all your debasement stuff. Now ur talking your bags again the opposite. Dummy

We are getting to that point of doom on Iran where X got to on Ukraine. Back then without gas and food we were going to be cholipping frozen, starved, dead Europeans of the streets. This time it's oil, gas, fertilizer and helium all about to bring about the end of humanity as we know it....

@Noahpinion Dude this is a really shit take haha.

However bad you think things are in the United States, they're worse in Canada and the UK.

@buccocapital Hahahahahahahaha

@packyM @bauhiniacapital Hahaha this isn’t just trolling, it’s bait

AI writing is taking over the internet. Every fundraising announcement, every X reply, every blog post. It's not just replacing human writing, it's overwhelming it. The problem isn't that AI writes poorly — it's that it writes plausibly. And plausible-but-empty is the most dangerous kind of noise. Here's why that matters: when my brain picks up on those subtle AI tells, I write off whatever you're trying to say — even if the idea itself was good.

@JC_ParetsX Nobodys talking about?! lol dude.

The relative strength has been in Hardware. New all-time highs today relative to S&Ps. No one is talking about it

@bubbleboi @TheGregYang Dude I really, really appreciate your write ups. This is a long form version of another tweet I was too dumb to understand. Thanks!!

Some other thoughts. 1. These guys had to know a lot of math. Like they are just applying it but you needed to really have an intuitive grasp of the machinery. I always liked random matrix theory and @TheGregYang was big into that so I thought idk was useful seems like lol it is. 2. You will probably eventually need dedicated encryption and decryption hardware to do this. Doing so is just going to make your token throughput way way way higher. Also it really does stress the importance of smaller data type multiplier like int4 becomes even more important I don’t think anyone was seriously using it for weights. 3. I don’t see KV shrinking. If anything you just unlocked way more context for free. Which means h100s are more valuable lol. The economics is really flipped on its head now cause like at 100k context KV uncompressed on Llama 70B is 32 GB you need multiple GPUs to inference that it won’t fit in an h100. But at 6x compression that’s 5.3 GB lol meaning it all fits on one h100. Nvidia is deff mad at this. 4. I see less HBM…. But really this is really bad for high bandwidth networking. KV cache exploding is why we needed 180+GB of HBM on a gpu but with tensor parallelism and pipeline parallelism you really just need to hold weights and if your KV is fucking 2.5 or 3.5 bits you also don’t need 800G SERDES. Sending 32GB at 800G takes 320 milliseconds not including overhead of FEC and lane alignment if you compress that to 5.3 GB that takes 53 milliseconds which is insane latency improvement. 5. There’s going to be more to come. I doubt this is the only way. Deepseeks technique of Multihead latent attention needed the model architecture and means training a new model . This is easy and you can do it to at INFERENCE time no model changes needed that’s huge.

I’ve been reading the TurboQuant paper a lot today and here are my thoughts/what I learned. One thing that stood out to me right away is how similar this is to a lattice based encryption scheme like LWE which I had experience implementing in hardware before for some glowies back in the day but in LWE all you do is take a vector you want to encrypt, multiply it by a random matrix, add some noise and the results look will look like uniform noise so you can’t decrypt it without a key. TurboQuant is relying on the same mathematical principles by multiplying the KV vector by a random orthogonal matrix. In geometric terms this is just scattering the values of your vectors across random axes on a high dimensional sphere. The reason this works in both cases though is because of to the way volume distributes in a high dimensional sphere. No matter what the original vector looked like, the projected values will always create the same shaped beta distribution. Because the beta distribution doesn’t have tails (it’s bounded [0,1]) and most of the values will clustered in the middle and this makes it perfect for quantization. Now if you have a known distribution now you can just quantize it by binning it into a histogram. There is an optimal binning algorithm for this that they used in the paper that is proven to create optimal bins for a given distribution (not just naively equidistant). They go with ~3 bits in the paper because that gives them 8 bins and because of the characteristics of the Beta distribution that’s enough to group things into meaningful clusters of information. The most interesting part of this to me (probably because I didn’t pay attention during LinAlg) is that they don’t need to decrypt to compute the attention score. Attention is just a dot product and that is preserved under any orthogonal transformation so you just rotate the query the same way you did keys and your lit you get the same numbers. They added another bit here for quantization error but yeah thats the gist…

@JessePeltan Terabase terafab or pick and place robots for ute scale. what's your bet works out best?

Almost looks like an array unfolding in space.

@Afinetheorem can yo upost the .md

Good writing is tough but I will say: I took a few recently published research papers, a .md file with very specific style notes (think New Yorker Editor), and Sonnet 4.6 + 5.4 Thinking, and asked for the introduction to be rewritten. They were *much* better. Try it yourself!

@duncancampbell Maybe I’m misreading your tweet, “disprove the baseload solar crowd” you mean the haters? I loved your whitepaper that basically said bess/ute solar could get to 90-95%. Who are you disproving? I’m confused

I’ve been building a very rigorous modeling stack, originally with the intent to disprove the “baseload solar” crowd. While I still think 100% uptime is foolish, I’ve become convinced 90-95% at reasonable LCOE is very doable. But it requires fully rethinking typical design.

Apparently a US diplomat has been carrying around a piece of electronics belonging to a foreign intelligence service Cannot make this shit up, we need jail time for these clowns

@KenBoessenkool Cope harder

Harper was judged as competent because of his government’s management of the global financial crisis. It was a consequence laden competence. Carney is being judged as competent because of his resume. It is a consequence free competence. +1

@Midnight_Captl I respect your AI and semi insights a ton. But respectfully, I think the long AI complex is factually correct, but gets overwhelmed near term and smoked pretty hard by sentiment catching up to recent events. ME flows next couple weeks prolly not gonna help

@CGasparino Clown take

Where was this brilliant trader screaming when oil hit above $100 after Biden invited Putin to invade Ukraine? Again the vast majority of stock and commodities traders are dumb; they trade off headlines. But you do have to ask yourself how we can’t get a peace dividend after taking out the lunatic mullahs?

@NateSilver538 @ramez Nate have you considered that perhaps the president is mentally unfit? Or is this war Biden fault too?

"Never start a bombing campaign that will completely halt shipping in the Strait of Hormuz" is right up there with "never get involved in a land war in Asia” in terms of the classic blunders.

@ShanuMathew93 30-40% real world RTE in all likelihood, endless issues scaling up manufacturing, limited if any actual real world production data outside of the lab? hmmmm

"A 30 gigawatt-hour iron-air battery from Form Energy is at the center of p Energy and Google’s Pine Island date center agreement, a scale that would make it the largest project announced by the energy storage company to date. Designed for 100 hours of discharge, the installation targets multi-day reliability events rather than the four-hour intraday demand shifting typical of lithium-ion systems."