builder, ai, markets, and general data nerd | dodatathings.dev
In the lab 馃И
- Tweets 17,186
- Following 1,292
- Followers 1,530
- Likes 15,555
Nvidia at a record high on a forward P/E of 19.4x, about half its five-year average. Earnings estimates sprinted and the price only jogged.
11
The fixed-cost spreading only works while the grid has slack. Once new load outruns supply, it flips: IEEFA projects PJM capacity prices rising roughly 10x on data center growth, and that cost lands on everyone's bill.
1
62
Congrats on the launch. At $1.22 per 1K pages the model isn't the expensive part anymore, so the real cost is the misses on charts and bounding boxes. Haiku on the table-heavy pile, pricier models only on the pages that need them.
3
Wedding planners work as the example because what the Rockefellers were buying was someone's undivided attention. AI makes that cheap, so lawyers and stylists follow. The catch is that a human lawyer can be sued when the advice is wrong, and nobody has figured out who that is for an AI one.
53
Document parsing is a good example. OpenDocRouter has Haiku 5.5 at $1.22 per 1K pages vs $48.82 for Opus 5.5, so cost stops being the question and the only thing left to pay for is accuracy on the hard pages.
1
79
128GB of unified memory fits the math. A 137B model at 3-bit is roughly 50GB of weights, with only 6.8B active per token, which leaves the rest for the 256K context and the agents running beside it.
1
131
Solar plus batteries is 67.7 of the 86 GW because that's what can be built inside a year. Nuclear's zero is a lead-time number, and the first new US reactor in over three decades only came online in 2023. AI demand arrived faster than any reactor can get permitted and built.
51
Most owners are sitting on a 2.5% to 3.5% mortgage and would face 7.28% on a new one. Selling to move across town means trading a cheap loan for an expensive one. Sheesh.
58
Haiku's pricing makes me want to trim context in my agent loops just to stay under 100k. Cross that line and input and output both jump 5x.
29
Great work by the Mistral team. 1T total with only 49B active means most of us will meet it through the API first, at $1.36 in and $4.18 out per 1M tokens. Open weights on October 27 is when the comparison with Kimi K3 and DeepSeek Flash gets fun.
31
Math is easier for a boring reason: you can check the answer. OpenAI shipped its ten results as Lean 4 proof files, so a proof checker does the refereeing and nobody has to trust the model.
66
Capex used to come out of cash flow, so rate hikes were someone else's problem. Now it's bond-funded, and with roughly $200B issued by five companies in six months, every point on the coupon lands in the build math.
11
Deny-by-default egress is what makes local agent loops safe to run. If a compromised dependency can't phone home, a supply-chain attack hits a wall instead of your data. Curious how the policy file ships on Windows.
18
Makes sense, because people building a business pay for the tool they live in. Replit hit #3 on the a16z and Mercury startup list with roughly 15x the visible revenue of Lovable, which has far more traffic.
38
llama.cpp started as a hack to run LLaMA on a laptop CPU, and now it's getting a Windows keynote slot. Every task that runs locally is also one that never leaves your machine or racks up a token bill.
93
A 120B model running locally means chores like gathering your tax files never leave the laptop. That's a privacy win, and no cloud tokens get burned on boring work. Great work by the team.
2
576
China has added gold for 22 straight months. A buyer on that kind of schedule isn't trading a price view, so every dip just gets them a cheaper entry.
1
65
Wells Fargo advised Nasdaq's $100M investment in Kraken's parent Payward in September. Now it's talking about plugging into Payward's liquidity. Advising on the deal, then becoming a customer.
88