Gabriel

@VibesPatrol @Londonmaxxing Where and what time is this? Already in shoreditch!

You know Londonmaxxing is real when you hit 2mil views on a post just about having a pint outdoors (standard post-work activities tbh) Come and join the maxis tomorrow at the second @Londonmaxxing event to recreate the meme irl with awesome people

@Andy_AJT Amazon here. Lmk if you wanted to check if its viable.

Planning the next Codex Community event... who's got great offices in London for a couple hundred agentic engineers to nerd out for an evening? Bonus points for outdoor space

@Andy_AJT Thanks. Requested to join Encode.

@gabrielchuv Best places are Encode and Shoreditch exchange, Encode for best vibes

Coworking in Shoreditch today Hmu if you want to join London sun + icecream + bank holiday = tokenmaxxing

@Andy_AJT When is the next event?

I’m now the OpenAI Codex Ambassador for London 🇬🇧 I am loving engineering with Codex & GPT-5.5 If you’re already using Codex or coding agents, building AI agents, or working on ambitious AI products in London I’d love to connect & see you at the next event!

@Agrippa_Inv @IREN_Ltd Agrippa you seem to have well researched takes. How does this acquisition make sense?

@IREN_Ltd $IREN is on a buying spree 💳

IREN has acquired Awaken, a creative and media agency specializing in content strategy and brand development for high-growth companies. Senior members of the team will join IREN, including Founder and CEO Chris Parker, who will lead IREN's brand and marketing strategy. Daniel Roberts, Co-Founder and Co-CEO of IREN, commented: “As we expand across new geographies and customer segments, brand awareness and customer engagement become increasingly important. Chris and the Awaken team have been trusted partners to IREN for some time, and bringing those capabilities in-house was a natural next step as the platform continues to scale." Learn more: iren.gcs-web.com/static-files/d…

@naval This is a problem in LLMs too. What do you think solved this? I think adversarial reasoning helps

The enemy of truth is motivated reasoning.

@litigious_dulce What about the 50mm drawdown clause? Is that enough of a positive sign if they only meet that within the clock?

NUAI: An Asymmetric Pre-Deal Position in AI Infrastructure Summary Investors comfortable holding $IREN, $WULF, $CIFR, or $APLD before they announced their first hyperscaler deals should be comfortable holding $NUAI today. The structural setup is materially identical with less execution risk. NUAI is operating the validated playbook those names established in 2025, with Stream Data Centers (Apollo-backed at $40B) and Macquarie already on the cap table, four hyperscalers as the only credible counterparties, and a six-month Macquarie clock functioning as a forcing function for lease execution. Position Overview Eighteen months ago, AI infrastructure names like IREN, WULF, APLD, and CIFR traded as speculative microcaps. Each re-rated sharply once a hyperscaler signed. Multiples expanded, floats compressed relative to opportunity, and the market repriced the companies from "miner" to "AI infrastructure platform." NUAI follows the same template. New Era Energy & Digital has 650 MW secured in Ector County, Texas — the flagship "TCDC" campus — and management has confirmed advanced commercial discussions with one of four hyperscalers: Alphabet, Amazon, Meta, or Microsoft. The joint venture was organized by the hyperscaler, who selected Stream Data Centers as development manager and an institutional capital partner (Northland believes Apollo) to provide equity and arrange approximately 80% project-level debt. Stream contributes hyperscaler relationships and operational execution. NUAI contributes site control. The structure was effectively delivered to the company. Why the IREN/WULF/APLD Comparison Holds The standard objection to any "early-stage X" pitch is that every microcap claims to be the next something. Four points distinguish NUAI from generic versions of that pitch: 1. Secured land. 650 MW in Ector County is owned outright, not optioned or under LOI. The recent equity raise eliminated the SharonAI overhang and consolidated full ownership of the TCDC site. Power-ready acreage is the binding constraint of the entire AI buildout and the single hardest piece to fabricate. 2. Institutional capital. Macquarie wrote a $290M project-level facility. Apollo acquired Stream Data Centers for $40B in November 2025 and is the implicit equity partner on TCDC. Both are among the most rigorous diligence shops in private capital, and both are staked. 3. Professional execution stack. Stream as developer/operator; RK Mission Critical for modular fabrication and supply chain; Thunderhead Energy for behind-the-meter power; Ramboll / EYP Mission Critical Facilities for engineering. Charles Nelson joined as President/COO in February 2026. Ted Warner — with nearly two decades of capital markets experience and over $7B in HPC-related financing — joined as CFO in March 2026. 4. Binary counterparty universe. Four hyperscalers, all investment grade, all capex-constrained on power, all publicly committed to multi-year buildouts. Whichever one signs represents top-tier credit on a 15-20 year colocation lease. Behind-the-Meter Has Become the Industry Default A year ago, the consensus view across the data center industry held that behind-the-meter (BTM) power solutions were unworkable at hyperscaler scale. Critics argued that hyperscalers required utility-grade reliability, regulatory complexity would prove insurmountable, and BTM would remain a niche workaround rather than a primary power strategy. That view was a real overhang on every developer pursuing BTM as a path to capacity. The consensus has reversed in twelve months. CIFR, APLD, WULF, and CORZ are all now executing BTM-led power strategies, and hyperscalers — facing multi-year interconnection queues and structural grid constraints — have endorsed BTM as a viable route to GW-scale capacity. Thunderhead Energy's role on the NUAI execution stack should be read in this context. NUAI is executing a strategy the industry has at this point publicly validated, with a power partner whose model is de-risked by parallel deployments at peer companies. This is a meaningful update to the underwriting. The power-delivery question that was an open risk on every pre-deal AI infrastructure name twelve months ago is now the operating assumption across the cohort. Stream Data Centers as the Execution Catalyst In November 2025, Apollo paid $40B for Stream — for a particular set of capabilities that map directly onto why a hyperscaler would select TCDC. Build-to-performance spec, not build-to-suit. Stream pre-aggregates standardized MEP equipment and configures it on the fly to customer specifications. The company quadrupled its development team during COVID and has been procuring long-lead equipment up to a year ahead of demand. Standardization speeds development time materially in a market characterized by acute power constraints and capacity scarcity. Configurable cooling that future-proofs the asset. Stream's proprietary cooling design supports air cooling and direct-liquid-cooling on the same footprint, scaling from 10-12 kW per rack to 400+ kW per rack. Customers can defer the air-vs-DLC decision until late in the build without extending the timeline, providing meaningful optionality across NVIDIA's roadmap from Blackwell to Rubin and beyond. Pre-existing hyperscaler relationship. This element has been broadly overlooked. Because Stream has worked with this hyperscaler before, we can safely assume that a significant amount of work product can be leveraged for TCDC. Management's fall 2026 lease execution target is credible because contracts are likely being adapted, not drafted from scratch. The distinction is between a startup negotiating with a hyperscaler from a blank page and the hyperscaler's preferred developer adapting an existing form to a new site. Execution risk lives in a different category. Expected Value Framework In my opinion, the probability of a deal with the current hyperscaler by August 2026 is 90%+. The hyperscaler organized the JV. They selected Stream. They directed the structuring. Engineering and permitting are progressing without observable friction. Negotiations leverage Stream's existing templates and shared counsel. The Macquarie facility requires lease execution within six months, aligning every party's incentives toward closing. As for the probability of any deal eventually, I would say 99%+. If the current hyperscaler exits — for which there is no observable reason in a market structurally short on power-ready supply — the structural work is already complete. Site control, partner ecosystem, financing template, and engineering package are not counterparty-specific. Another publicly traded data center company recently demonstrated this dynamic: a hyperscaler counterparty exited, a replacement was secured, and the timeline extended by approximately one month. Stress-tested at a deeply conservative 50% probability of a deal — well below what the structural setup supports: 50% × 4-5x upside ≈ 2.0-2.5x expected return 50% × 50% drawdown ≈ 0.25x expected loss Net expected value: approximately 1.75-2.25x At 90% probability, expected value approaches 3.5-4x. The asymmetry is wide enough that halving the upside and doubling the downside still produces a positive expected value. Re-Rating Mechanics: Why a Deal Drives 200%+ From Here, Not 10% A market-microstructure point underlies the upside case. When mature AI infrastructure names — IREN, WULF, CIFR, APLD at current scale — announce hyperscaler deals, the stock typically moves around 10%. Optionality is already embedded, and announcements function as confirmation rather than revelation. Smaller, less-followed names behave differently. DGXX has announced materially smaller deals than what NUAI is contemplating and moved 50%+. Expectations are not embedded, the float is small, and the announcement forces a re-rating from speculative microcap to credible AI infrastructure platform (Note that a deal cannot be priced in because many institutions are waiting to buy until after a deal is announced). NUAI sits closer to the $DGXX end on market cap and visibility but closer to the IREN/WULF/APLD end on asset quality and counterparty caliber. That mismatch is the opportunity. A first hyperscaler deal at TCDC could plausibly drive an immediate 200%+ re-rating — not because steady-state fundamentals support that exact multiple, but because microcaps gap rather than incrementally re-price. Investors do not get to scale into the new range. Downside is bounded by the existing balance sheet, which is clean post-Macquarie and post-equity raise with no SharonAI overhang. Upside is a non-linear re-rating event. The Case for Data Center Exposure A reasonable question, given the breadth of the AI investable universe — semis, photonics, custom silicon, robotics, model labs — is why allocate to data center developers at all. Data center economics are durable in a way most AI-adjacent verticals are not. Hyperscaler colocation leases run 15-20 years. Counterparties are investment grade. Cash flows are recurring. Once a campus is leased, it produces something close to a bond. EQIX has compounded through every macro cycle of the past fifteen years on this dynamic, and the structural reason is simple: an AWS region does not get turned off because the economy slows. Compute demand is structurally inelastic at the margin, and existing infrastructure is locked into multi-decade obligations. The asset class is also tractable for non-specialists. Underwriting reduces to power, land, customers, and contract terms. Many other AI-adjacent verticals — photonics, custom silicon, neuromorphic, edge inference — are genuinely interesting and likely lucrative, but the underlying technology evolves quickly enough that most investors cannot reliably assess winners. Data centers fit Buffett's "in pile" — comprehensible, durable, and underwritable on standard metrics. The constraint is that asymmetric opportunities within the data center space are increasingly scarce. For WULF to 5x from current levels would require multiple gigawatts of new capacity, additional contracts, and substantial revenue growth — achievable but grinding. NUAI requires one announcement with one of four hyperscalers for Phase 1 of TCDC. The bull case condenses to a single press release. For investors who participated in the 2025 IREN/WULF/HUT/APLD/CIFR cycle, NUAI offers the same trade structure with two improvements: the underlying thesis has been validated by the prior cohort's outcomes, and the macro evidence — exponential capex guides, tightening power constraints, structural undersupply — is materially stronger today than it was eighteen months ago. Conclusion NUAI is structurally identical to the IREN, WULF, and APLD trades in early-to-mid 2025, with three improvements. The thesis has been validated by the 2025 cohort's outcomes. The execution stack — Stream / Apollo / Macquarie / Ramboll on day one — is more institutional than what several of those names had at first announcement. And the forcing functions are tighter, with a six-month Macquarie clock combined with a hyperscaler-organized JV on 650 MW of secured Texas power. The position reduces to a single proposition: one press release reprices the equity by triple digits. Downside is bounded by an institutional cap table and a clean post-raise balance sheet. The expected value math holds at 50% probability and compounds at the 90%+ probability the structural setup supports. Simply put, this is a remix of the IREN/WULF/APLD trade.

@karpathy @simonw Why does it make sense for oai or other labs to not pursue this?

Yeah exactly. It’s such a cool concept for a product. It doesn’t seem like oai will continue pushing the direction, (which makes sense) but I hope a startup can clone it and actually give it care, iteration and make it work and imo a lot of people would really love it. More generally, the product roadmap of big labs is clear and predictable, which also leaves big pockets of opportunity for startups, one of the biggest ones is this I think.

I think it's non-obvious to many people that the OpenAI voice mode runs on a much older, much weaker model - it feels like the AI that you can talk to should be the smartest AI but it really isn't

@poetengineer__ @threejs I like this. I feel the learning with AI field is moving into a few interrelated directions that will or could merge at some point. Knowledge accumulation (karpathys work), learning base (helping with structure) and a model of user understanding (which I haven’t seen much of)

one direction from this that excites me: a learning base instead of a storage one: not for what you already know, but for what you don't. made one for deep reading of plato's timaeus. 2 things i carried over: non-rag, indexed fs, and /raw-is-sacred to separate sources from generated content. a few features i find genuinely helpful:

@alexanderrX_ @lombardo187 @aruni_t Nah you can easily rent a decent 1 bed on less than that without compromising your finances. I do. More than that, you don’t need to be too long on even half that amount to be able to buy

@lombardo187 @aruni_t Read my reply again.

london is one of the few cities on earth where someone earning £250k a year is renting a one bed flat or living with flatmates. this city humbles everyone equally.

@gabriel1 I think this is true but there’s an opportunity to build a UX to guide people into the rabbit hole

historically the hardest part of learning is knowing what is relevant to learn & finding the information. it's so hard we decided to put everyone through 20+ years of general learning with chat models this is not true anymore, you can deterministically learn anything top down

@gabriel1 This is clear to me too. An AI mentor should be able to understand what you understand and update that as you learn. GPT already has some awareness of the state of your understanding and tries to consider that when responding. Do you think this advance will come from the AI labs?

i could learn any topic in 5 minutes with the mist optimal text & visual explanation, if it's adapting live to what you do and dont understand fundamentally ai has like another 5 gpt3 to gpt4 moments ahead

@tenobrus Why langchain?

gigafucked: - grammarly - calendly - miro - retool - webflow - langchain - writer - harvey - glean - expedia - monday fucked: - accenture - intuit - notion - jasper - canva - alphasense - postman - airtable - talkdesk - sierra - zapier - replit - solace probably fucked: - cursor - pilot - clay - mercor naively seems fucked but so competent / plugged in they seem to be figuring it out on the fly anyway: - linear

@_trish_xD The other 10% is realizing somebody else’s stupidity

i swear 90% of debugging is just realizing your own stupidity in slow motion

@loganthorneloe Define context engineering?

I just sat in an internal presentation + demo that blew my mind. Research speed will 100x in the next few years. Context engineering is king.

@gabriel1 Doing new things. Novelty is part of the reason why as a child one feels like everything goes slower

it seems like doing dumb shit and traveling works, i went spontaneously to mexico and it felt like i a full week although it was just a weekend. we just did constant out of distribution ideas

what are your best ideas for how to slow perception of time? if you want it to feel like you are 7 years old again and a 1h lecture feels like a day, how do i slow down time to that?

@ben_m_somers Why? Not disagreeing, just curious

I'm less interested in AGI than AHI Augmented Human Intelligence

@patro_satyam AI mock coding interviewer algomentor.co.uk

lets drive some traffic to your site. what are you working on this weekend?

@Anubhavhing AI mock coding interviewer algomentor.co.uk

Pitch your startup in 3 words.