Wednesday Morning

The action has shifted to post-training: reasoning boosts, RLVR, better eval harnesses, and tighter data and rollout quality. Those are iterative and observable, so they get replicated. Lambert's take is that this locks in a fast-follower equilibrium unless the training paradigm changes, closed labs gain a durable edge from proprietary user data, or open builders hit funding limits. Fully open recipes like OLMo 3 are already lagging without access to those levers.

The scale keeps creeping up either way. Tracker snapshots moved from 762,099 BTC to 766,970 BTC, roughly $54.8B, putting Strategy around 3.6%-3.7% of Bitcoin's 21M supply. At that size, each teaser isn't just sentiment anymore, it's read as timing for one of the biggest recurring bids in the market.

An oil field near Erbil is reported to have exploded and caught fire, with details still thin but the signal getting louder. Two separate posts about six minutes apart described the same event, including one placing it on the road to Erbil and citing Naya. No confirmed cause, casualties, or damage yet, just a burning site hitting energy infrastructure in a region already seeing repeated blasts.

The second wave is full workflows. Claude Design + Opus 4.7 gets wired into a single personal dashboard across calendar, tasks, finances, and health, while Claude Code runs as a terminal builder with memory, reusable skills, and MCP connections to Slack, GitHub, and Drive. One demo had Opus 4.7 clipping, editing, scheduling, and posting videos continuously.

It's live as a research preview on Mac for Pro users, still early and token-heavy. The same rollout also saw a brief Codex outage that lasted about 10 minutes; the system helped fix itself, but the key insight came from a human, and rate limits were reset after recovery.

Alongside that, Gemini Enterprise is becoming a full agent control plane, not just model access. It now bundles an Agent Gateway, governance and scaling tools, and 90+ partner-built agents from Adobe, Atlassian, Box, Oracle, Salesforce, ServiceNow, Workday, and others, pushing Google Cloud toward a full agent ops stack with its own runtime and distribution layer.

Martin shared swarm-forge, a tmux-based control system for coding agents that lets multiple agents work side by side, assign tasks, use separate git worktrees and merge changes under human supervision. He said the setup can handle large coding batches on its own, including writing hundreds of tests in a single run, generating Gherkin specs, refactoring and running mutation-style test work.

It is also being framed as open source and something to build on. There is a public repo and app server docs. The direction is agents moving from demos into real execution across files, terminals, and desktop apps.

The shift is tying value to three demand loops: gas from tokenized assets settling on Ethereum, ETH as core DeFi collateral, and staking yield. About a third of supply is already staked, and BlackRock's staked ETHB ETF is being used to frame ETH as a yield-bearing reserve asset rather than software.

So a model roughly 15x smaller in total parameters is posting better coding scores than Alibaba's prior flagship, while dropping hardware needs from ~807GB class setups to something in the ~56GB range (or lower with quantization). Same general capability tier, but suddenly in reach for local runs instead of datacenter-scale setups.

Pipecat released version 1.0 and introduced Subagents v0.1.0, expanding the project from voice agents into a framework for realtime, multimodal, multi-model AI applications. The software arrives after two years of development, with Pipecat saying it is already used by thousands of startups, scale-ups and enterprises, and the team used a multiplayer game called Gradient Bang to demonstrate parallel agents sharing context over an event bus.

That's a roughly 3x jump in stated model size across what's framed as back-to-back monthly releases, not a slow quarterly cycle. The real shift isn't just bigger models, it's tying the very next major version, Grok 5, directly to AGI. Two scale-ups, then the finish line, instead of an open-ended roadmap.

The new push is handling the stuff that usually breaks these setups. Deck says it works through MFA and CAPTCHA and adapts when the UI changes, instead of retrying brittle selectors. That targets the main failure point in no-API automation and makes these flows usable inside tools like Claude Code or Cursor without constant babysitting.

This reads like the moment energy inflation stops sitting at the pump and starts moving through the supply chain. Diesel was up 88% year over year, and that hits everything that moves food around before it even becomes a shelf price. Fertilizer, plastics, and other inputs haven't fully landed yet, so this is still early-stage. The cost base for food manufacturers is already shifting higher.

The pushback shifts blame onto domestic policy instead, arguing net zero has raised bills and left the UK exposed to imported energy shocks. Kemi Badenoch is tying that to a specific counteroffer: a "Cheap Power Plan" she says would cut household bills by £200 by removing VAT, scrapping green taxes, and expanding North Sea drilling.

The shift isn't one agent helping with a task, it's one person running 20-30 agents in parallel and acting like a manager. In software, that means dozens of branches moving at once, with agents writing code, iterating, and responding over multi-hour loops instead of a single IDE session. Howie Liu's framing: one agent on a branch can match about three engineers at roughly 3x speed, turning strong ICs into supervisors of stacked, parallel output.

A "third path" for AI oversight is edging toward reality: independent auditors checking safety claims instead of companies grading themselves or governments running everything. The setup is state-licensed but market-based, similar to financial audits or product safety checks. Gillian Hadfield tied this directly to Dean Ball's recent push, and Virginia has already ordered a formal study after bipartisan votes of 84-14 and 40-0.

Paying with stablecoins just got pushed onto existing retail hardware instead of crypto-specific setups. WalletConnect Pay is now live on iMin Android POS terminals, enabling USDC checkout across more than 100 countries, with support for any wallet and supported stablecoin. The shift is that merchants don't need new devices, it runs on deployed POS, moving crypto payments directly to the checkout counter.

The key is reuse: adjacent tokens share about 38% of experts, around 3-6× higher than an independence baseline, and that held across 13 tasks and 7 languages. That overlap keeps routing costs from exploding and lets speculation stack gains. It also finds a non-monotonic speed curve with a sweet spot at moderate batch sizes, plus a batch-size-1 bonus from amortizing fixed overhead.

Kalshi is being put far ahead on fees, but the numbers are running into a transparency problem. Artemis pegs **$5M daily fees** on average, about 4.25x Polymarket and 3.2x Hyperliquid, with a 7-day estimate around $5.1M versus $1.2M and $1.6M. That makes Kalshi look dominant in the current prediction market race. The pushback is about what sits under those totals. Polymarket activity is publicly auditable. Kalshi does not show who generates its volume, where liquidity comes from, or how much is driven by market makers. The lead exists on paper, but the inputs behind it are not visible, so the comparison is being treated as incomplete.

At the same time, basic fuel delivery is still a bottleneck: more than 160 Alaska communities get just one or two shipments a year, usually by barge, and price swings are hitting both carriers and customers while putting delivery timing at risk. That pressure was raised directly with Interior as part of the latest push.

A single $100M BTC-backed loan just hit Maple's books, showing large borrowers are still taking size under its stricter, fully overcollateralized setup. This isn't retail flow, it's institutional credit with offchain custody and liquidity parked in stablecoins and T-bills. Maple's run so far: $15B+ issued over 2+ years, $100M+ in interest paid out, and zero losses.

The banking layer runs through Erebor Bank, N.A. (FDIC member), while the stablecoin side is explicitly not FDIC-insured and can lose value. Compliance, routing, and reconciliation are built in, so treasury and payments teams can move between fiat and stablecoins without stitching systems together.

Bundling payments, social, mobility, and services into one daily-use layer is where the value piles up, and the claim is Web3 should copy that instead of shipping single-purpose apps. The shift is from individual-first products to community-driven infrastructure with built-in economic activity, then extending it into tokenization, stablecoins, and AI coordination. Animoca is already running 130+ autonomous agents on Minds, framing orchestration as ecosystem plumbing, not a feature.

More meme-coin chatter is showing up alongside people actually rotating into higher-risk tokens again, after a stretch where that behavior cooled off. OSF pointed out it's the first time in a while both the content mix and positioning appetite are shifting together, calling it a possible vibe change. In crypto, that combo usually lines up with retail moving off the sidelines and back into speculative trades.

"Non-custodial" is getting treated less like a design choice and more like a red flag for illegality, collapsing tools that never hold user funds into the same bucket as bad actors. Janusz G's point is that this shift is driving selective enforcement: legitimate self-custody infrastructure gets targeted, while online casinos built on regulatory arbitrage keep operating with far less pressure.

Most LLM evals dodge real finance work, especially outside English. EDINET-Bench goes straight at it with ~41,000 Japanese securities reports and scores models on fraud detection, earnings forecasting, and industry prediction. It just landed at ICLR 2026, and the paper is out, pushing evaluation toward messy, domain-specific filings instead of clean English benchmarks that miss how this stuff actually works.