←
Jesus Rodriguez16,380 posts
Jesus Rodriguez avatar✓

Jesus Rodriguez ✓

@jrdothoughts

Co-Founder, CTO-CPO of @SentoraHQ (fmr IntoTheBlock), Co-Founder of @layerlens_ai, @neuralfabric( acq by Cisco) and The Sequence, Teaching at Columbia-Wharton

◷ Joined September 2011▧ 327 Photos and videos
Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·19h

The latest Sequence Knowledge, Issue #973: "Learning RSI: World Models as Training Grounds for Self Improving Agents." Agents learn by practice, and practice needs worlds. Someone has to build the warehouse before an agent can learn to navigate it. World models change that: a learned simulator generates the consequences of an agent's actions, and paired with a system that proposes tasks and grades attempts, the training environment becomes part of the improvement machinery. DeepMind's SIMA 2 and Genie 3 are an early demo. Gemini proposes tasks and rates attempts, and training on that self-generated experience improved most held-out tasks. That's transfer within generated worlds, not robots yet. The catch: convincing worlds can teach the wrong lessons, reward models can confuse "near the goal" with "done," and generating experience costs real compute. The real question: do skills learned in generated worlds survive outside them? thesequence.substack.com/p/t…

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 16

DeFi composability gets interesting when developers can turn it into a product people actually use. At @SentoraHQ, you can now integrate Dynamic with every vault on vaults.sentora.com. That brings wallet infrastructure and onchain yield closer together. The workflow: select a vault, choose Dynamic and your framework, and generate integration code preconfigured with the vault’s address, chain, and deposit token. Use the generated widget where supported, or connect the vault module to your own interface. Architecturally, the pieces fit together: Dynamic provides authentication and embedded wallets; Sentora’s integration provides access to vault positions, deposits, and withdrawals. Users stay inside your application and sign transactions with their own wallets. For a fintech building an Earn feature, that creates a practical path from familiar onboarding to onchain allocation. For developers, it reduces the work of stitching together wallet connections, contract interactions, and vault configuration. What excites me is the distribution potential. A consistent integration workflow across vaults makes it easier for application teams to bring different yield strategies into their products. We’re building toward a world where adding onchain yield becomes a normal software capability: composable, accessible through familiar interfaces, and connected directly to the applications where users already manage their money.

CyHro - CEO of Sentora, Firelight avatar
⟳ Jesus Rodriguez retweeted
CyHro - CEO of Sentora, Firelight@admff492·Nov 15

Kraken BTC Earn typically increases in yield as BTC price drops. This is because the yield is in USD, and the BTC notional drops in dollar terms. There is always an opportunity in market chaos.

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 15

Most of the coverage of Jensen Huang stepping down is about succession, the stock, and whether NVIDIA can keep shipping without him. Fine questions. But I think the more interesting one is about everyone else: what happens to an AI market that was partly designed and financed around one man's vision when he leaves? There's a nice story about the AI boom in which thousands of entrepreneurs independently discover that the future requires NVIDIA chips. There's also a slightly more complicated one, in which NVIDIA invests in the entrepreneurs, guarantees some of their revenue, and helps arrange the money they use to buy the chips. The invisible hand, it turns out, has excellent investor relations. That's the subject of the latest Sequence Radar, "The Invisible Hand Wore a Leather Jacket": thesequence.substack.com/p/t… The logic was elegant. If you don't love your customers' bargaining power, finance more customers. CoreWeave, Nebius, Nscale, open models, nearly $50B into frontier labs, minimum-revenue guarantees that let lenders underwrite neocloud capacity. NVIDIA got to be supplier, shareholder and backstop at once. A strategic investor earns three returns on one check: equity, hardware sales and a healthier ecosystem. A normal investor gets one. Which means some of these businesses were worth financing to NVIDIA specifically, and arguably to Jensen specifically. Nobody wrote down a "Jensen put." But believing someone has reasons to rescue you does a lot of work in credit markets. Finance has built entire neighborhoods on less. The $500B+ third-party financing push hadn't produced a deal even before he left, and a successor may want each check to justify itself. The perverse outcome: the world keeps buying more AI while the market consolidates into exactly the hyperscalers Jensen's ecosystem was built to counterbalance.

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 13

The internet has spent years asking bots to identify traffic lights. Give them purchasing power, and suddenly everyone would like to schedule a partnership meeting. ChatGPT Wallet is a massive moment for agentic commerce because it connects AI’s distribution to finance’s existing infrastructure. openai.com/index/introducing… OpenAI is working with Visa, Mastercard and American Express to let agents pay using tokens that protect actual card details. Every purchase still requires user approval and a security check. The initial rollout is limited, but ChatGPT already reaches 1.2 billion people weekly. The future can arrive on your regular credit card statement. That is a powerful adoption strategy. My bet: businesses will increasingly compete to become an agent’s preferred supplier. Accurate product information, dependable fulfillment and systems agents can navigate become distribution advantages. The customer still decides what matters; software increasingly handles how to get it. OpenAI envisions purchases delegated within defined limits as a future step. Imagine that model extending across travel, procurement and everyday spending. The bigger shift is an economy reorganizing around software acting on human intent. Payments give everyone a financial reason to participate. Apparently, the fastest way to make businesses take AI agents seriously is to let them bring customers’ money.

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 12

A neocloud can have eager customers, busy GPUs, and an uncomfortable call with its lender, all in the same week. A financing squeeze can arrive before a demand slump. In the latest Sequence Opinion I look at why. thesequence.substack.com/p/t… There are four pressures. Money goes out before it comes in: sites, power, and GPUs get paid for before customers pay. Delivery is harder than signing, because a cluster only bills once the grid connection, permits, and cooling are done. GPUs have an uncertain economic shelf life. And "diversified" demand often leads back to the same few wallets: CoreWeave's top three customers were 72% of its Q2 revenue. Take a cluster that collects $100 and pays $40 to operate and $50 in debt service. If collections drop to $85, a $10 cushion becomes a $5 shortfall, even though the cluster is still busy. My view is that this changes ownership and business models before it proves AI infrastructure was a mistake. Buyers will pick off power connections, sites, and contracts rather than whole companies. The telecom buildout is the closest parallel: society got the networks, but the original investors didn't always get their money back. The question I keep coming back to: is delivered capacity turning into cash fast enough to cover the debt?

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 11

The first answer is what gets the applause. The useful part is what happens next: whether the model can recover, revise, and stay inside the constraints it was given. In the latest Sequence Learning Loop I look at three releases through that lens: GPT-6 Cyber, Inkling 2 Small, and Gemini Omni 1.2 Flash. thesequence.substack.com/p/t… GPT-6 Cyber is trained against a blunt signal: a working exploit, or not. On ExploitGym it reaches 48.5%, up from 32.1% for GPT-5.6 Cyber. A Linux pipeline reproduced 1,724 vulnerabilities and 948 privilege-escalation exploits, and in a honeypot test it stayed in scope 99.93% of the time. Inkling 2 Small is built for work that continues across many actions, trained on more than 17 million rollouts, including screenshot checks for web apps. If each of 100 steps is 99% correct, only about 37% of full runs succeed. Longer context alone cannot fix that. Gemini Omni 1.2 Flash extends clips to 15 seconds and lets you edit them conversationally, at $0.07 a second in 720p. The valuable part is changing the lighting without undoing a decision you already approved; a beautiful new clip can still be a failed edit. The question I keep coming back to: what is the quality of the second attempt?

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 10

Give a language model a hard theorem and enough time, and it can produce a remarkable amount of mathematics. The troublesome part is deciding which of it deserves to survive. An elegant paragraph can conceal an invalid inference. A correct final answer can emerge from a broken argument. Train on either mistake and the next model repeats it more convincingly. That's why math is such a useful lab for recursive improvement. A prover generates candidates. A checker decides whether they establish a precise statement. The checking rules stay fixed, so the feedback can remain informative even as the generator gets stronger. In the latest Sequence Knowledge I look at three versions of this loop: AlphaProof, Goedel-Prover-V2, and DeepSeekMath-V2. thesequence.substack.com/p/t… A formal kernel gives you a trustworthy signal, at the cost of formalization. A learned verifier is more flexible, and also fallible. It can predict the right score while inventing the wrong explanation. And even a perfect checker only validates the theorem it was given. Change the statement and the certificate becomes empty. The question I keep coming back to: does generation N produce experience that makes generation N+1 a better prover, or just a more fluent one?

nikos avatar
⟳ Jesus Rodriguez retweeted
nikos@definikos·Nov 9

If you're curating credit for frontier industries, @jrdothoughts Financing Prometheus is worth the read. He draws the line that matters: representing an existing security, making it useful inside DeFi and originating new productive credit are three different jobs. Putting a bond onchain does not finance a compute facility that still needs to be built. From a curator seat the hard part is clocks and first loss. Onchain lending as we run it today is short-duration, overcollateralized, and liquidator-dependent. A multi-year construction loan against customer contracts is long-duration, often thin on spare collateral and recovery runs through courts and replacement operators. Those are not the same product. Aave Horizon already shows the compliance trade: permissioned RWA collateral, nontransferable receipts, allowlisted participants. That preserves legal wrappers. It also removes secondary exit for liquidators. Where vault structure fits: - senior / junior tranching so somebody with a balance sheet owns first loss - async vaults (ERC-7540 style) and redemption queues sized to actual loan cash flows, not to a utilization curve - exit buffers so depositor T_exit does not force a fire sale of illiquid claims - NAV oracles with bounds, not a 24/7 mark that pretends a private credit note trades like ETH - legal wrappers and servicers who can enforce, because code cannot seize a half-built data center Jesus is right that secondary transfer can change the investor while project funding stays in place. That is the only way patient capital works onchain. But transferability without a willing buyer and a junior buffer is just a queue with a coupon. Tranching reallocates losses but it does not fix bad project economics. Illustrative: $100m facility, 3-year build, $15m junior first-loss, senior sitting in a curated vault with T+30 redemptions. senior_loss ≈ max(0, ΔNAV × notional − junior_buffer) A 20% NAV write-down on incomplete construction: 0.20 × 100 − 15 = $5m into senior. If the vault's exit buffer is thinner than the redemption queue that day, depositors eat duration they did not underwrite. So the number to watch is junior buffer vs the worst NAV path before redemptions clear. x.com/jrdothoughts/status/21…

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 9

A couple of weeks ago, @SentoraHQ and @aave announced their partnership for building a liquidity hub on Aave v4. We just crossed $100M in deposits and we have a super exciting roadmap.

Post attachment
Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 9

we are rapidly learning that the biggest blocker for super intelligence might not be compute or energy but the readiness of political institutions. The irony....

Jesus Rodriguez avatar
Jesus Rodriguez@jrdothoughts·Nov 9

I just published an essay, Financing Prometheus, on the future of DeFi in frontier industries and “after crypto”: jrodthoughts.medium.com/fina… “After crypto” is not an exit from digital assets. It names an expanding economic purpose. Crypto remains part of the system, while repayment increasingly comes from electricity delivered, computing capacity used, goods shipped, and discoveries made useful. The assets of that future are the claims that make this possible: equipment and facilities for intelligence, the power that sustains them, the receivables that keep commerce moving, and the milestones that carry science from a laboratory into production. These are not substitutes for tokenized Treasuries or fund shares. Representing an existing security, making it useful inside DeFi, and originating new productive credit are distinct achievements. The frontier requires all three. A company building a computing facility may have valuable customer contracts and very little spare collateral. Somebody must evaluate those contracts, finance construction, allocate losses, and commit for an appropriate term. Putting an existing bond onchain does not perform those jobs. DeFi’s opportunity is to organize confidence for that next industrial age: capital markets that can underwrite uncertainty, assemble entire capital structures, and give promising enterprises time to become productive. We have no shortage of extraordinary possibilities. The work now is to put foundations under them.