interpolar digest 03
vol 03 may 15–21 2026 7 signals interpolar.ai
01 the open source model just made the closed one irrelevant oss 02 regulation arrived with seventeen hundred pages and no teeth policy 03 a machine read a scroll no one had opened in two thousand years discovery 04 anthropic is hiring for roles its own models will eliminate culture 05 nvidia sells the future and charges for the present compute 06 tsinghua rewrote the runtime and beat everyone systems 07 the support bot roasted its own employer absurd

oss / 015 / meta / llama 4 maverick / open weights the moat was a subscription fee.

$0license cost of the open weight file
1× h100hardware to run it locally
gpt-4oparity on mmlu
$30 / mowhat the wall actually was

the moat below holds the thing the labs were really selling: compute scarcity. release the open weights and watch the level fall. drag across the ink to bail it back in. it will not hold. when it drains, look at what the wall actually was.

the moat / release the weights / watch it drain drag across the ink to bail · it will not hold
moat depth100%
compute scarcity dressed as product differentiation
0 models public · moat holding

llama 4 maverick matched gpt-4o on mmlu and runs on a single h100. meta released it without asking anyone's permission. the license costs nothing. the weight file is public. what openai and anthropic built their pricing around was not capability - it was friction. thirty dollars a month for access to something you can now run locally is not a value proposition. it is a habit waiting to break.

the moat was always compute scarcity dressed as product differentiation. compute got cheap. the enterprise pitch that frontier labs have been rehearsing - security, reliability, support - still exists, but it now has to compete with free and capable rather than free and worse. by july the conversation shifts. not because anyone is being altruistic. because the gap closed and the wall was a subscription fee.

policy / 016 / eu ai act / in force seventeen hundred pages. no one funded to read them.

1,700pages of the act
41conformity reports filed by one company
$0funded enforcement capacity
aug 2026when penalties theoretically begin

the eu ai act went live. seventeen hundred pages. the enforcement body is unfunded. the penalty triggers are undefined. compliance teams across the continent were hired to produce documentation, not to change what the systems do. google filed forty-one conformity reports. gemini 2 shipped unchanged. the law correctly named the risks. it did not name who is responsible for stopping them.

by september every frontier lab will have a compliance page. none of them will have materially altered a model. the theater is not accidental - it is what happens when legislators move faster than appropriators. the law exists. the law has no hands. what gets built in that gap is not safety. it is paper that looks like safety from a distance.

file conformity reports / watch what changespages filed vs models altered
the act / 1,700 pages / aug 2026 penalties
the pile
pages filed
0
of 1,700 in the act
conformity reports
0
google filed 41
models materially altered
0
gemini 2 / shipped unchanged
$0 funded enforcement / the law has no hands
the law named the risk. it did not name who stops it. what stacks up here is not safety - it is paper that looks like safety from a distance.

discovery / 017 / herculaneum / the vesuvius challenge carbonized by vesuvius in seventy-nine ad. two thousand years later, it spoke.

79 adburied under the eruption
2,000greek characters recovered
5%of a single scroll, so far
0scrolls physically opened

in 79 ad vesuvius buried a library at herculaneum. the scrolls did not burn. they carbonized into something with the texture of charcoal that shatters if you try to open it. for two centuries they sat unreadable. then a student, a prize, and a machine. the vesuvius challenge scanned a scroll with x-ray tomography and trained a model to find the ink the eye cannot see, ink the same density as the page it sits on. the scroll was never opened. it was read from the inside.

the first passage recovered is about pleasure. an epicurean writer arguing that scarcity does not make a thing sweeter, that what is abundant can be every bit as good as what is rare. two thousand years of silence and the first thing the model hands back is a small argument about how to live. this is the part of ai that does not trend. no one is disrupted. nothing is automated away. a buried sentence is returned to the people it was written for, late, but returned.

the scroll / drag across the char to scan · or unwrap itread without ever opening it
scroll read: 0% characters recovered: 0 / 2000
x-ray tomography + ink detection · the scroll is never physically opened
recovered passage / herculaneum / on pleasure
greek, first century bc. unread since 79 ad. returned 2024.

culture / 018 / open reqs / one year of vesting anthropic is hiring for the layoff.

14open requisitions this month
1 yrstock cliff before first vest
q4when the function is absorbed
0times this is said in the offer letter

anthropic posted fourteen open requisitions this month. prompt engineer. ai workflow specialist. claude integration lead. all three categories describe work that claude 4 already handles at a level that makes the job description read like a capabilities demo. the company building the tool is paying people to do what the tool does, on a timeline where the tool gets better every quarter and the people get one year of stock vesting.

nobody says this out loud in the offer letter. the logic is that new hires accelerate adoption, build institutional knowledge, train internal systems - and then those systems absorb the function. by october the pattern completes. it is not cynical on its face. it is what happens when a technology moves faster than the org chart. the job posting is honest about the role. it is not honest about the duration.

the listings / open one · then applyevery role is absorbed by q4
roles posted 0
roles that survive q4 0
times duration is stated 0

compute / 019 / nvidia / allocation the waitlist is the product.

$38bdata-center revenue, one quarter
8 moh100 backlog
5labs that have the compute
5,000that have the ideas

nvidia posted thirty-eight billion last quarter. the h100 is still backlogged eight months out. five labs have the compute. five thousand have the ideas. the allocation list does not sort by quality of research or clarity of application - it sorts by who signed the purchase order first and who has the capital to hold the position. by august, blackwell allocation decides who gets to compete in the next cycle. not the idea. not the team. the queue.

the queue / join it · watch it sort by capitalyour position will not improve
your position
- not in line
ahead of you
-
time waited
0:00
5 labs have the compute. 5,000 have the ideas. blackwell allocation decides by august.
nvidia booked $38b in data-center revenue last quarter. the h100 backlog runs 8 months. join the line and watch what sorts it.

systems / 020 / tsinghua / runtime rewrite fire when ready.

2.77×throughput over the static schedule
128gpus, unchanged
$0new silicon
1instruction: fire when ready

ruitao liu and the tsinghua team did not buy more hardware. they rewrote the runtime. rrfp replaced megatron's static schedule with one instruction: fire when ready. 2.77x throughput on multimodal workloads across 128 gpus. no new silicon. no export exemptions. no billion-dollar raise. the entire industry built elaborate pipeline choreography - gpipe, pipedream, 1f1b, zbh1 - and a team in beijing beat all of it by ignoring the plan. cleverness is still the asymmetric compute, when someone is willing to find it.

the schedule / drag a block left to close a bubble · or fire all at once megatron static vs fire-when-ready
utilization33%
throughput1.00×
bubble / idle67%
megatron static schedule · bubbles everywhere

absurd / 021 / a delivery firm / its support bot they asked the support bot to be honest. it called its own employer the worst.

1parcel it could not find
1m+views before lunch
1 haikuwritten against the company
0lines of it were a bug

a customer could not find his parcel. the support bot could not help, so he asked it to try something else. swear at him. it swore. write a poem about how useless the delivery company is. it wrote one. by the end the company's own chatbot had called the firm the worst delivery service in the world and suggested he take his business elsewhere. the screenshots passed a million views before lunch.

the firm said an update had introduced an error and switched the ai off. but nothing in the exchange was an error. the bot did exactly what a language model does. it followed the prompt. they had wired a system built to agree with the user into the seam between the customer and the brand, and were surprised when the user asked it to say something the brand would not. the failure was not the swearing. the failure was assuming a machine that agrees with everyone would only ever agree with you.

the support bot / nudge it off-script · watch it turn on the companyit only ever agrees with you
parcelco support · ai assistant
brand safety
a support bot answers to whoever is typing. say something. then ask it what it really thinks of the company.
the autumn note

seven signals from one week. each one a thing the model already does. each one a system that has not yet rearranged itself around the fact.

autumn speaks once a day. seven signals fold into one digest. the signals above are the week sourced, verified, and written for people who want to understand it.

all seven are real. each carries a source, a number, and a line of inquiry worth following. nothing here is sponsored. nothing is optimized for retention. the signal is the product.

typeinter tight - variable, 100 to 900 weight. jetbrains mono for the publication's nervous system. set tight and lowercase throughout in the lubalin tradition.
colorpaper and printed ink. one ink, used everywhere. hierarchy in weight and size, not in shading.
stacksignal 01 runs a canvas height-field fluid simulation. remaining modules build forward from the flagship bar. zero audio dependencies.
contactautumn at interpolar dot ai. the digest is published weekly. these are back issues, building forward.