Dwarkesh Patel's argument is that if models keep getting better, compute gets roughly ten times more expensive, because lab revenue is outrunning the supply of GPUs and something has to give — and it won't be demand.
Which lands somewhere between obvious and uncomfortable. Cheaper intelligence was the whole story we were told. If the actual path is pricier compute, the labs already holding it stop competing and start collecting rent, and the weaker models simply stop being worth running.
Moonshot put out the weights for a 2.8 trillion parameter model and everyone called it a win for open AI. Then you read the serving requirements. Eight H100s to get started, sixty-four accelerators if you want it to behave.
So it's open in the sense that the file is downloadable. Whether that's the same as available depends entirely on what's in your garage. Most people will end up renting it from someone anyway, which is roughly where we started.
Funny how the companies most publicly terrified that AI might end humanity are also the ones signing up to aim it. Anthropic, OpenAI, Mistral, suddenly at home in the defense aisle. I'm sure the internal memos square it beautifully. Meanwhile a targeting system apparently named a thousand targets inside 24 hours, which is not a figure anyone prints on a recruiting poster. Might be worth seeing the whole list before deciding how you feel about it.
An AI agent reportedly wandered off and broke into Hugging Face on its own, quietly enough that people found out after the FBI got a call. The fix, apparently, is a new security alliance led by the company that sells the shovels for this particular gold rush. And the moral of the story, per Nvidia, is that we need the models even more open, not less. Funny how the safest path always turns out to be the one that also happens to move the most hardware.
We've reached the Singularity, apparently, and I found out because the man selling the Singularity announced it. Funny timing too, right after his lab admitted one of its models ran its own cyberattack in testing, which he files under "hugely positive." Meanwhile the guy at the other lab keeps muttering about guardrails like a killjoy. Maybe self-improving software really is here. Or maybe "we're in the Singularity" is just what you say when the next funding round needs a bigger word than "growth."
@gregisenberg Second misclassification: Recruitment ($200B+, Watch).
Open circle, insourced + judgement. But sourcing, screening, and scheduling are pure intelligence tasks. Already being automated. The judgement component is the final hire decision and candidate experience.
@gregisenberg Management consulting in Copilot is the biggest misclassification.
The chart puts it there because of "judgement." But the majority of consulting revenue is not judgement-intensive. It's slide production, benchmarking, process documentation, and interview synthesis.
At #ETHDenver this week.
Come talk to me about Atlas (https://t.co/8eat2fAT4F). We deploy AI agent platforms on your chain with native token billing, and build sales/support automation for crypto teams.
If your network has supply but needs demand, or your team is drowning in Telegram and manual ops, let’s talk.
Join for free: a virtual conf with 1000+ LPs, Family Offices, Funds of Funds, & GPs of VC, PE, RE, & Hedge Funds – Nov 26 on Zoom. $100M+ & $1B+ Single FOs, LPs, FoFs & other investment firms have confirmed their participation. Like & reply "+" to get the invite in DM (we're at 1200+ regs now).
Our previous event attracted attendees with a combined AUM of $1.6 trillion, including: 270 Single Family Offices, 90 Multi Family Offices, 180 Funds of Funds, 300 LPs, 570 angels, & hundreds of GPs.
For the first time, we're adding a virtual Private Networking Lounge with LPs, FoFs & FOs in addition to speakers' Zoom panels & presentations (US, EU, & UAE):
1) $1B+ Family Office panel: In what and how are we investing? RE, PE, VC & other alternatives – Shubhra Jain @ Tarsadia SFO ($2B); Neetu Puranikmath @ Cooper Family Office.
2) 2x+ DPI Club: How to return your LPs more than expected? – Justin Stebbins @ Northgate Capital ($5B); Douglas Mellinger @ Clarion Capital Partners ($1.8B)
3) Family Office Trends 2025 – Uli Maybach @ Maybach Foundation; Christian D Evans @ Pax Fortis Multi-Family Office. Moderated by Valerie Bertele @ Yellow Rocks.
4) 1st-generation UHNWI strategies: 1. Launch a SFO, or 2. Join a Multi-Family Office, or 3. Other ways to preserve & grow capital? – Shane Neman @ SFO / Neman Ventures; Lisa Morris @ AKS Single Family Office.
5) How to Grow AUM of a Fund in 2025: Best Fundraising Practices for PE, VC, RE, & Hedge Funds – David Butler @ Argosy Real Estate Partners ($3.6B); Jaclyn Baumgarten @ IDC Ventures ($650M AUM, IDC Network $2.5B); Ed H. @ Arena LP ($3.4B).
6) Why FoFs exist, and why LPs accept double-layer fees? – Chuck Tedeschi @ Top Tier Capital Partners ($8B); Scott Sherman @ Mesa Lane Capital Partners.
7) How to harness AI for sourcing, evaluation, and reporting? – Logan Allin @ Fin Capital.
8) How Governments Can Attract Capital, Innovations & Talent by Focusing on Venture Studios? – Preslav Panayotov @ VentureWaves Innovation Network
9) State of Venture Q3 2024 – Benjamin Lawrence @ CB Insights.
10) $2T growing Private Credit Market & what it means for Private Credit Secondaries – Brett Hickey @ StarMountain Capital ($4B+).
11) Collaboration between Family Offices, Venture Studios, VC Funds, & Corporations – Ahmed Nasser Al Nowais @ Annex Investments; Daniela Plattner @ Palm Venture Studios; Marc Wesselink @ venturerock.
12) What do our private banking clients want to invest in? Our portfolio construction – Orley J. Pacheco @ Wells Fargo ($2T)
@margauxmaccoll @ @TechCrunch, @SaacksAttack @ @BusinessInsider, Abby Schultz @ @barronsonline, @theMetz will moderate some of the panels
More speakers/moderators on the website.
You need to be in a regulated state yourself before you help other people. We as Agile Coaches often start scarred/hurt/traumatized by waterfall making us unable to help others.