Chandra. K

building anthroly. we help companies sell their data to ai labs and build rl environments + datasets with 100+ phds. posting what i see inside ai data deals

[Chandra. K @sk_2496]

i run anthroly. we do two things

help companies sell the data they're already sitting on to ai labs. support threads, call notes, expert docs, the stuff nobody looks at

build rl environments and datasets for labs. 100+ phds on contract, 400 more ready to go

i post what i actually see inside ai data deals. follow if that's your world

[Chandra. K @sk_2496]

the most valuable thing your company owns might be the stuff you were about to delete

@Google pays @Reddit $60m a year for old forum posts. labs pay phds up to $200 an hour to write answers

and most founders are sitting on years of support tickets, call notes and expert docs they call junk

what's the oldest folder you still have?

[Chandra. K @sk_2496]

@NandoDF followed you 🔥. would love to keep talking to you about various topics

[Chandra. K @sk_2496]

@h7software not yet. a phd and a grader arguing over one rubric line for way too long would honestly make decent stream content

[Chandra. K @sk_2496]

@jasonjias @X building datasets and rl environments for ai labs with a bench of phds. always-on agents are the thing i'd most want real eval data on. how do you test yours?

[Chandra. K @sk_2496]

@MaximLevoshin helping companies sell their data to ai labs, and building datasets for labs with phds. games and edtech teams usually sit on more useful data than they think

[Chandra. K @sk_2496]

@Viswana34226652 where are these companies based ? currently ai labs are buying USA based company data only.

[Chandra. K @sk_2496]

the cumulative clause is the sleeper. a lot of real post-training now is a small team, a pile of expert-written tasks and an rl environment

calling that frontier misses where the compute actually is

[David Sacks @DavidSacks]

Is 10²⁵ FLOPs Still the Frontier?

A new Anthropic-written safety bill is circulating on Capitol Hill. Some of its ideas could be reasonable if limited to true frontier companies. As written, however, it will eventually sweep in most of Silicon Valley, damaging the startup ecosystem.

The bill defines the “frontier” as any model trained on 10²⁵ FLOPs — the EU standard from 2023. Biden’s executive order, California’s SB 53, and New York’s RAISE Act used 10²⁶ (ten times more). 10²⁷ is where the frontier actually is today (e.g., GPT-6 Astra).

Astra was reportedly trained on 100,000+ Grace Blackwell GPUs. A 10²⁵ model can be trained on as few as several hundred GB300 chips in three months. That is not frontier in 2026.

The definition is also cumulative: it adds FLOPs from fine-tuning, reinforcement learning, or other material modifications. In practice, that captures almost all meaningful post-training — work that is becoming commonplace in the enterprise.

The enterprise trend is toward thousands of smaller custom models built for specific workloads. These pose no frontier risk. Treating them as frontier systems would impose large compliance costs across corporate America for no safety gain.

Dario has claimed that Anthropic designs rules to disadvantage frontier companies like his own while advantaging smaller competitors. This definition does the opposite.

A stale 2023 cutoff favors incumbents who already clear the bar and can absorb the costs. It hobbles the open model ecosystem that threatens Anthropic’s pricing power.

The frontier keeps moving. Any definition meant to track risk should move with it.

Anthropic should clarify if they really believe 10²⁵ FLOPs represents the frontier.

[Chandra. K @sk_2496]

@jwt0625 does "smooth mapping" end up meaning you pick training systems by symmetry rather than by volume? feels like the opposite of how llm data gets picked

[Chandra. K @sk_2496]

@waseem_s the raters shape it as much as the spec does. character training is basically a preference dataset with opinions, and whoever grades it leaves fingerprints

[Chandra. K @sk_2496]

@OliverMolander and the unglamorous bottleneck inside post-training is tasks a grader can actually check. compute is easier to buy than that

[Chandra. K @sk_2496]

@HarryStebbings price matters most when the moat in the deck is data anyone can rescrape next quarter. that's the diligence question i'd ask first

[Chandra. K @sk_2496]

@levie same thing is happening with training data. everyone assumed labs would stop buying once the internet was scraped. demand for expert-written data went up instead

[Chandra. K @sk_2496]

@NandoDF that cost is real. the middle ground i keep coming back to is grading a small sample of full traces properly and using those to audit the cheap verifier, not replace it. self-distillation with nobody auditing is where the bias compounds

[Chandra. K @sk_2496]

@ebelilov @VAIBHAV22155287 rl post-training for diffusion feels underrated. what's the reward here, a learned preference model or something you can actually verify? that choice seems to decide everything downstream

[Chandra. K @sk_2496]

@_AxelT anything that says "this looks off" before i hit send. spell check did it long before ai did. what's the one mistake maple tracker is built to catch?

[Chandra. K @sk_2496]

@SoritesResearch "fuel worth more than the engine" is going in my notes. that's the whole data licensing market in one line

[Chandra. K @sk_2496]

@levie the bottleneck you land on is almost always review. agents can start ten things, someone still has to decide which one was right, and nobody budgeted time for that part

[Chandra. K @sk_2496]

@satish_vutukuru 19 rows down to 11 is the most useful note here. the rows that disappear tell you more than the new ones

we're in a similar lane at anthroly, 100+ phds on contract building rl envs and datasets for labs. always up to compare notes on what labs actually ask for

[Chandra. K @sk_2496]

@DavidSacks curious how they'd even measure cumulative flops for a company fine-tuning someone else's open model