←
amargoel980 posts
amargoel avatar

amargoel

@amargoel

@bitoHQ CEO -- context-driven model routing. I like to build, and learn. First startup did $30M in revs, next did $50M, then went public $250M - @PubMatic

◷ Joined May 2008▧ 37 Photos and videos
amargoel avatar
amargoel@amargoel·Nov 8

If one lab's AI speeds up its research faster than everyone else's, its lead grows on its own. That's the story of how to use AI to create a real lead. For a company like Anthropic heading to an IPO, that could be pretty valuable. I wish we could see more of what METR saw.

amargoel avatar
amargoel@amargoel·Nov 8

But Anthropic had the contractual right to redact, and it removed most of the evidence METR found most useful. A check on how it splits compute between capability and safety work was dropped midway. Anthropic says it's protecting secrets from competitors.

amargoel avatar
amargoel@amargoel·Nov 8

METR spent three months inside Anthropic. Its finding: Anthropic's AI is now speeding up Anthropic's own model improvement research by about 2x, and roughly 80% of that speed-up comes from AI itself. Across all frontier labs it's closer to 1.4x.

amargoel avatar
amargoel@amargoel·Nov 6

The winners from here are whoever already has data centers running. If you build software on top, watch token prices. I'm no longer sure they keep falling as fast.

amargoel avatar
amargoel@amargoel·Nov 6

The limit on AI used to be chips. Now it's permits and power. If you can't build, the compute that already exists gets pricier. Anthropic's S-1 showed it this week: buying short-term compute took Q3 from profit to a $567M operating loss.

amargoel avatar
amargoel@amargoel·Nov 6

Wednesday: voters pass most of 41 local measures to slow data center construction. Amazon +6%, Alphabet +3%. Less capex, more cash. Thursday: Nasdaq -5.5%, worst day since April 2025. Nvidia -9%. AMD and Micron -17%. Big Tech down too. Looks random. I don't think it is.

amargoel avatar
amargoel@amargoel·Oct 28

5/ One more thing that stood out: over 80% of last quarter's revenue came from models Anthropic released in the previous six months. Customers move to the newest model almost as soon as it ships.

amargoel avatar
amargoel@amargoel·Oct 28

4/ But it's getting less concentrated, fast. In Q1, Anthropic's 5 biggest customers were 36% of its revenue. By Q3 they were 21%. The rest of the customer base is growing faster than the giants.

amargoel avatar
amargoel@amargoel·Oct 28

3/ The filing doesn't say how much revenue each group brings in, so I estimated it. My rough math: the 250 biggest customers bring in a bit over half of revenue, and all 11,900 together bring in about three quarters. So yes, it's top-heavy.

amargoel avatar
amargoel@amargoel·Oct 28

2/ Anthropic has about 40 million paying customers. Of those, 11,900 pay it more than $100K a year, 1,900 pay more than $1M, and about 250 pay more than $10M.

amargoel avatar
amargoel@amargoel·Oct 28

1/ Anthropic filed to go public this week, and its IPO filing (the S-1) is the first real look at the business. I read it. One claim I keep seeing is that Anthropic really only sells to a few hundred companies. So I checked.

amargoel avatar
amargoel@amargoel·Oct 21

A lab's hiring plan is its most honest capability forecast. They are setting up a fund to pay the people they are no longer hiring.

amargoel avatar
amargoel@amargoel·Oct 21

Anthropic is pledging 5% of its equity to cushion AI job disruption (CNBC, Oct 18). On the same weekend, an analysis of 1,680 of its engineering resumes showed junior technical hires down from 27% to 9%.

amargoel avatar
amargoel@amargoel·Oct 19

Meta published the weights for Muse Spark Open this week: 816B params (42B active), top US open-weight model per Artificial Analysis at a 39 Intelligence Index. The global board: GLM-5.2 scores 51, DeepSeek V4 Pro is at 44, Kimi K2.6 43. America's best open model trails China's third best. But America is making progress! It's still a real gift if you want to code privately on your own hardware (Meta demoed it on two Mac Studios). The license is not quite open though: gated download with your legal ID, mandatory attribution, a $20M revenue clause on API resale.

amargoel avatar
amargoel@amargoel·Oct 18

DeepSeek released V4.1 Flash 1017 yesterday. We ran it through our extensive, long-term, long-agent coding benchmarks: same tasks, same prompts, same grading we use for every model. The cheap model just reached Sonnet's level. On long agentic coding sessions it scored 77%, matching Claude Sonnet 5.5 and Haiku 5.5 at their defaults, for $0.14 a session against Sonnet's $1.28. On quick one-shot coding questions it tied Claude Opus 5.5 outright at 90%, for about 1/25th of the cost. The gains over the previous DeepSeek sit in analysis and planning, where its plans now score level with Opus. It also got steadier: not one derailed session in 16 runs, where the old version occasionally answered the wrong thing entirely. It still falls short where mistakes are expensive: implementing and auditing code changes, where it trails Opus by 12 to 17 points. That gap hasn't moved between versions. If we were setting this up for a team, we'd point DeepSeek at the investigation and planning work and keep Opus on anything that ends up in production.

amargoel avatar
amargoel@amargoel·Oct 12

Google's new Gemini "coworker" agents get real Workspace accounts: their own email, calendar, and a spot in the company directory. Two days later, Satya Nadella wrote that companies should treat AI models as insider risks. Real accounts do help: scope, permissions, logs, rate limits. But right now, access is shipping much faster than the guardrails. Buckle up! It could get crazy.