Pinned Tweet
we're cooking something big, stay tuned ;)
With all the discussion about AI’s progress this week, I keep coming back to how we govern what these systems actually do.
When an answer or action depends on an agreement, you should be able to see the rules behind it.
Here’s what I mean.
“Can I stop a prepayment?”
The answer might take you through three agreements. You should be able to follow the reasoning and see the language behind it.
If we use agreements to coordinate economic activity, why would we expect AI agents to operate without understanding them like this?
Understand · 1 of 4
Nvidia funds OpenAI, OpenAI pays Oracle, Oracle pays Nvidia. It works great while demand grows, but if it slows, one company's revenue is literally the other company's loan payment.
@AnishDAC spent '08 inside CLOs and CDS, so he knows what happens when things go off the rails – at Sunol, CA
These are genuinely nice charts, @a16z. But you skipped circular financing, so I made a quick video on it (see below 👇).
A little about me and why this is relevant. The weekend before Lehman failed, I was on the advisory team that helped issue their final CLO. I spent the whole great financial crisis (GFC) in CDOs, CLOs, and CDS. What I took away from that period of time was EVERYTHING runs through the agreements, and those agreements are all interrelated.
So when Oracle’s 5yr CDS hit a record last week, wider than 08, I decided it was time to pay attention. We know debt informs equity, and when protection gets that expensive, you really have to start asking what if?
So let's follow the money:
→ Nvidia funds OpenAI
→ OpenAI buys compute from Oracle
→ Oracle buys chips from Nvidia
But what if demand drifts to cheaper open models (a very real concern at the moment)?
• The labs would grow slower than the projections they raised the debt on
• But they’d still owe the same
• Debt Service Coverage Ratio takes a hit
• Each company’s revenue is another one’s bill, so the whole circle of participants eventually feels it
Btw, you only really see this coming with certainty by understanding the agreements and how they work together. Normally, this takes weeks with intense iteration. But you know what? The market won’t give you weeks when it really matters... The movie Margin Call had it right: be first, be smarter, or cheat.
I’d rather be the first two, never the latter. This is one reason I built DAC: so anyone can SEE and understand the risks and opportunities through scenarios embedded across agreements, before you need them, or instantly if required.
It is very important to know your options before the trade unwinds because when it unwinds... it unwinds fast!
Now, I am not saying this will happen, but it is important to know what-ifs so one can plan and be positioned to react accordingly
sarick retweeted
One Xerox bond issue (yup, they're still in business, believe it or not) lost 13.4% of its market value in two weeks. Roughly $45m.
Here’s why knowing your agreements matters to your money.
Xerox’s $500m of second-lien notes due 2031 traded from 67 on Feb 4 to 58 on Feb 18, 2026.
During that period, Xerox moved IP into a joint venture and raised $450m. Existing debt documents weren’t amended.
Ropes & Gray explained the opening: the venture fell outside the definition of “subsidiary,” so key protections aimed at subsidiaries didn’t reach it.
The permissions were already there.
And the refinancing pressure is building. Octus’s September data shows North American term loan B and high-yield maturities of:
• 2028: $597bn
• 2029: $646bn
• Combined: more than $1.2tn
That gives stressed borrowers more reason to explore LMEs.
Do you know whether your agreements let them:
• Move assets out of lenders’ reach?
• Put new debt ahead of yours?
• Waive protections without your consent?
Those vulnerabilities and protections matter to your price, leverage and recovery.
So does the institutional memory: what was negotiated, why, and what changed. Too much of it sits in spreadsheets, emails, Word docs and counsel’s head.
You need that knowledge available before someone acts on a path you didn’t see.
The video traces Xerox through eight documents and 24 connected steps.
Know your options before you need them. 👇
sarick retweeted
hyperscalers, lenders, private credit: do you know what your agreements commit you to, what protects you and when your options change?
if a tenant goes bust, who has to step in? who can decline? what do they owe if they do?
DAC’s what-if lets you test that before it happens. change the assumptions, see the available paths and trace the answer back to the clauses.
quick illustration below using a public Google / Cipher / Fluidstack agreement.
know your options before you need them. 👇
Called it – at Sunol, CA
Anthropic's discovery of array-associated reverse transcriptases?
Turns out it was already done by a Danish scientist. He used Claude for administrative tasks, but not in an agentic manner for the discovery. He already filed a patent and gave talks on it
nytimes.com/2026/09/27/scien…
Would love a unification layer for all these agent harnesses...
Muse is great, but I don't trust Meta with sensitive data. Spark is insanely good at YouTube flows and pretty limited outside that. Grokbot actually feels like a control plane for company ops. Instinct was promising until it got painfully slow.
Right now you end up picking a different harness for every job. I want one interface that lets me use the best parts of each without inheriting all of their trust assumptions.
Maybe A2A can make a comeback – at Sunol, CA
sarick retweeted
. @tobi, we’re building DAC to help agentic commerce scale with liability controls.
We turn agreements into rules that control what agents can actually do. The buyer’s authorization and the merchant’s terms govern the transaction before either agent acts.
Easier to show. Here’s Muse getting its headphones. 🎧
Happy to walk your team through it.
an llm reading a loan agreement has to infer structure, resolve cross-references, track state, and explain itself in one pass. that's why you can't audit it.
we compile the agreement into a graph first. reasoning is traversal, not inference.
There’s a lot of concern right now about AI acting against our interests.
Agreements set the guardrails. Sometimes those guardrails leave room for actions we don’t want.
“Can they move our collateral without our consent?”
In this example, yes. The lender’s agent checks the rules and flags the risk while the transfer is still pending.
You should be able to see those guardrails, understand what they allow and monitor what happens next.
That shouldn’t stay buried in the fine print.
Monitor · 2 of 4
a lot of the world’s knowledge never makes it into software because it dies somewhere between thought and typing.
google adding voice-first drafting across docs, gmail, and keep is part of a bigger shift.
work software can finally start absorbing the messy context people actually carry: tone, memory, corrections, half-formed ideas, lived expertise.
including from people who know what they mean, but might not write like an office worker.
inc.com/ben-sherry/google-wi…
The biggest voice AI markets are also the least measured.
Hundreds of millions of users speak languages where training data is scarce, dialects are underrepresented, and benchmarks do not reflect how people actually speak.
That is the gap we have been working on at @psdnai.
We have been collecting real-world audio data at scale across low-resource languages, noisy environments, and production conditions.
sonar.psdn.ai is an evaluation framework that makes that work useful to the broader field.
Collecting data is only half the loop. The other half is knowing where the model fails across datasets, speaker groups, acoustic conditions, and error types, then using that signal to collect the data that actually improves it.
SONAR makes that measurable. This is a major step toward voice AI that actually works outside the narrow distribution of clean, high-resource English.
Voice AI has an evaluation problem. Models look strong on public benchmarks, then collapse on real-world audio.
Introducing sonar.psdn.ai: a recipe-driven evaluation framework for low-resource languages, real-world audio, and production failure modes.
Details ↓
system-level dictation is a very practical preview of where voice AI is going if it gets the right training data.
less typing, fewer app-specific workflows, and more software that understands intent from natural speech.
voice and physical AI should disappear into daily life this way - useful, broadly accessible, and materially better.
elevenlabs has made it trivially easy to deploy a voice agent across 70+ languages without a dev team.
the interesting signal around this is the speed of adoption. when the tooling gets this accessible, the competitive gap between companies that move and companies that wait compresses really fast.
your support stack is either getting smarter this quarter or falling behind.
experts are finally converging around the fact that the internet is tapped out.
synthetic data as the proposed fix can’t fully offset real-data depletion, since you can’t solve a quality problem by generating more of the thing that’s already degrading.
the only answer is data that was never online to begin with, but out in the real world.
this now becomes a logistics and incentives problem - who shows up, what they capture, and how it gets processed and licensed.
Yeah, you know what they say -- data being the "new oil." And apparently, like oil, it's a finite resource.
(My latest in Forbes) via @forbes forbes.com/sites/joemckendri…
the next wave of AI is physical - robots, voice agents, AVs. but that won’t happen without the right high-quality data, especially once you move into high-stakes deployments.
numo brings that effort to life, starting with bengali, hindi, tamil, and telugu.
to get the volume of data needed, you have to build the supply, task by task and market by market.
real-time voice translation just shipped as a product. zoom. teams. mobile. api.
now the cost of serving a customer in their language has collapsed. not “we’ll add spanish support in Q3.” now. the companies that figure out global GTM before their competitors do are going to look very smart in 18 months.
language was always a distribution problem dressed up as a hiring problem. that excuse is gone.
techcrunch.com/2026/04/16/de…
the new yorker altman piece is long but worth the read.
the part that stuck with me was the structural question underneath all the drama. the most capable AI systems in the world are being built on gulf sovereign capital, embedded in military infra, and racing toward a trillion dollar IPO. every government watching this is now asking the same question: do we want our critical AI infrastructure controlled by entities with competing interests?
that’s why the race for sovereign AI is accelerating. and why who contributes to and controls the data powering these systems really matters.
.@RonanFarrow and @AndrewMarantz interviewed more than a hundred people with firsthand knowledge of how Sam Altman, the head of OpenAI, conducts business. They also obtained closely guarded documents that have not been previously disclosed.
newyorker.com/magazine/2026/…
this piece on japan’s physical AI push has a signal buried under the surface - labor shortages are forcing physical AI deployment timelines that the data infra isn’t ready for.
once physical AI is load-bearing, the data requirements change completely. you can’t train reliable warehouse or factory systems on generic web data. you need environment-specific, long-tail examples from the exact conditions these robots operate in.
that data doesn’t exist in public datasets. it has to be collected, licensed, and structured from scratch.
techcrunch.com/2026/04/05/ja…
the thumbnail says it all:
all the craft is in curating the perfect dataset.
@Microsoft, please let me introduce you to real-world training data for physical ai.
semafor.com/article/04/08/20…
