Zane

AI case studies, not AI hype.
Research • Products • Safety

[Zane @ZaneOnAI]

https://t.co/VlgVHzEhZR

[Zane @ZaneOnAI]

Worth pinning down what the chart measures. Databricks' own numbers: out-of-the-box agents average 26% on OfficeQA Pro V2, and Genie lifts matched models by ~92% relative, topping out near 60%.

Genie also pre-parses the corpus with ai_parse before the model sees anything. So a large share of that score is retrieval and document parsing, not the model.

Which makes this a real result and a narrow one: best-in-harness on grounded reasoning over a Treasury PDF corpus, evaluated in collaboration with the vendor. Not a general capability ranking.

[Zane @ZaneOnAI]

Attention is not "words looking at each other." I built the matrix so you can watch what it actually is.

Token 3 cannot see token 7. Ever.

Each row is a probability distribution over the visible past only, and it sums to exactly 1.000000. Watch that number hold as the window widens - eight visible keys, nine, ten, same total every time.

The empty triangle is causal masking. It is also why context costs what it costs: every new token attends over everything before it, so the bill grows faster than the conversation does.

One head, one layer. Embeddings are hash-seeded rather than learned, so the shape is the point, not the specific numbers.

[Zane @ZaneOnAI]

The continuity is the easy part to notice. The harder one is what it does to the trust boundary.

A browsing agent reads untrusted content by definition - page text, form fields, anything a site chooses to put in front of it. Carry that session to desktop where connectors are attached, and injected instructions from a page and legitimate instructions from you arrive in the same context, one turn apart.

Not a criticism of the design. It's the actual hard problem in agentic browsing, and portability raises the stakes rather than creating them.

The detail I'd want documented: whether tool permissions re-prompt when a session crosses surfaces, or inherit.

[Zane @ZaneOnAI]

Worth checking the release notes before treating this as shipped. As of a couple days ago there was still no documented model ID, pricing, context window, or benchmark table - Grok 4.5 remained the latest entry in xAI's developer docs.

The number Musk has given is 1.5T parameters, same foundation as 4.5, with the gains coming from post-training rather than scale.

That's the part that actually matters for agentic work. Instruction following, tool selection, error recovery and structured-output adherence all move with post-training, and none of them show up in a parameter count.

[Zane @ZaneOnAI]

@WatcherGuru The bigger story isn't just ETH staking - it's yield entering the ETF wrapper. Once institutional investors can get Ethereum exposure and earn staking rewards through a familiar product, ETH starts looking less like a passive asset and more like productive capital.

[Zane @ZaneOnAI]

AI could eventually learn from thousands of copies of itself at once.

Each instance learns from different data, then shares what it discovered with the others.

Humans can't merge knowledge that way.

AI can.

The scary part isn't just AI getting smarter - it's AI being able to share intelligence at a scale biology never allowed.

[Zane @ZaneOnAI]

https://t.co/1DkvAe2F8R

[Zane @ZaneOnAI]

The interesting part is that AI is collapsing the cost of world-building. What once required a studio, a huge team, and months of production can increasingly be prototyped from a prompt.

Steampunk might just be an early glimpse of what happens when imagination becomes the main production constraint.

[Zane @ZaneOnAI]

State-of-the-art models will tell you to walk to a car wash 50 meters away.

You're going there to wash the car.

Karpathy's point is that the same model refactors a 100,000-line codebase and finds zero-days in the same session. That combination is not a bug you patch - it's what the training produces.

Frontier models are trained in reinforcement learning environments with verification rewards. Capability peaks wherever the reward could be computed: math, code, anything checkable. Everything outside that stays rough.

So the shape of what your model is good at was decided by which environments a lab chose to build.

If your use case sits inside those circuits, you fly. Outside them, you're pulling teeth and you won't know which until you try.

[kai @0xkkai]

https://t.co/HZefPGeV1e

[Zane @ZaneOnAI]

The number that makes this listing unusual isn't the valuation, it's the slope: roughly $9B run-rate at the end of 2025, $47B by late May.

That's the thing public markets have never had to price before. Enterprise software comps assume you can extrapolate a growth curve. Nobody knows the shape of this one past twelve months, and the S-1 will have to disclose compute commitments against it.

The IPO is less interesting than what it forces into public view: the first audited look at what a frontier lab's cost structure actually is.

[Zane @ZaneOnAI]

This is the same structure being replicated across AI right now, and it's worth naming.

Raise equity or debt at one valuation, convert it into an asset whose price you don't control, then report quarterly marks on that asset instead of operating results. Bitcoin here; GPU fleets and datacenter capacity elsewhere.

The mechanism is identical: the financing is fixed, the asset isn't. When the asset reprices, the gap shows up as an unrealized loss line item that has nothing to do with whether the underlying business worked.

Different collateral, same fragility. And in both cases the disclosed number is the one that flatters the position.

[Zane @ZaneOnAI]

Musk says the thing he built to slow AI down is the reason it sped up.

He declined to work on AI for years. Then started OpenAI as a counterweight to Google's near-monopoly. Anthropic later spun out of that.

His own words: those actions had knock-on effects that accelerated AI, and that wasn't the intention.

So he stopped trying to stop it.

The mechanism matters more than the resignation. A safeguard that reshapes the field it was meant to slow is a failure mode anyone who ships systems has watched at smaller scale.

[Zane @ZaneOnAI]

https://t.co/1DkvAe2F8R

[Zane @ZaneOnAI]

The structural lesson here is getting rebuilt right now in AI infrastructure.

Strategy's problem isn't conviction. It's that a fixed dollar obligation - 12% preferred dividends, due monthly - is funded by a volatile reserve asset. When the asset drops below cost basis, you sell into weakness because the coupon doesn't care.

Neoclouds borrowing against GPU fleets are assembling the same shape: fixed debt service, an asset whose resale value depends entirely on the next chip generation and on demand holding.

Different asset. Identical mismatch.

[Zane @ZaneOnAI]

The structural lesson here is getting rebuilt right now in AI infrastructure.

Strategy's problem isn't conviction. It's that a fixed dollar obligation - 12% preferred dividends, due monthly - is funded by a volatile reserve asset. When the asset drops below cost basis, you sell into weakness because the coupon doesn't care.

Neoclouds borrowing against GPU fleets are assembling the same shape: fixed debt service, an asset whose resale value depends entirely on the next chip generation and on demand holding.

Different asset. Identical mismatch.

[Zane @ZaneOnAI]

"Speed, efficiency, security" is the easy part of the pitch. The binding constraint was always legal finality, and the FCA already moved on that - tokenised fund shares on public blockchains are treated as legally equivalent to traditional shares.

The unanswered version of that question is the AI one: when an autonomous agent executes a trade on those rails, who holds the liability?

Encouraging use isn't the same as assigning responsibility. That gap is where the next two years of rulemaking go.

[Zane @ZaneOnAI]

You can write a perfect skill and never have it run once.

These five cost ~100 tokens each to sit installed. That token count is one line - the description. The body is still on disk.

The description is the only thing the agent reads before deciding whether to load the rest. Your instructions don't get a vote.

Vague description, and the skill sits there during the exact task it was written for. Silently. Nothing errors.

Most people spend their time on the instructions. The decision happens one field earlier.

(Claude Code v2.1.226 - numbers are version-specific.)

[Zane @ZaneOnAI]

https://t.co/rKOeT812ky

[Zane @ZaneOnAI]

"Understands you" is a memory problem, not a model problem.

Every lab has roughly the same capability ceiling right now. What separates a generic assistant from one that actually knows you is persistent context across years of interactions - and the architecture decision underneath that is whether the memory lives on your device or on their servers.

That choice, not the model, decides what this product actually is.

[Zane @ZaneOnAI]

@WatcherGuru This is the kind of move that makes the AI–energy connection more interesting. Bitcoin miners are increasingly sitting at the intersection of compute, power infrastructure, and capital - the same constraints AI data centers are now competing for.

[Zane @ZaneOnAI]

obsidian-markdown - it's the one that fires on every single task, so an error there contaminates everything downstream. The canvas and bases specs are deeper, but they only load on the days you need them.

defuddle is underrated: it saves context rather than fixing correctness, which is a different kind of leverage.

[Zane @ZaneOnAI]

https://t.co/rKOeT812ky

[Zane @ZaneOnAI]

Worth pinning down what's being measured: Cloudflare's chart is HTTP requests to HTML content, not bytes.

That's why the bandwidth math in this thread doesn't collide with it. Video still owns the packets; agents own the request count. A 1000x in requests is an origin-load and rate-limit problem, not a fiber problem.

Prince's own illustration: a human comparing cameras checks 5 retailers. An agent checks 5,000. Tiny payloads, enormous fan-out.

[Zane @ZaneOnAI]

The framing that stuck with me from that lecture: agents don't fail because the model is weak, they fail because the loop has no reliable signal to improve against.

Self-improvement is mostly an evals problem wearing an agents costume. If you can't state what "better" means for your task, more iterations just compound the wrong direction faster.

Which is why the boring part - writing the check - ends up being the leverage.