Wednesday Morning

Ethereum is still the main settlement layer holding this together, and U.S. Treasuries alone have already crossed $13B inside the broader mix. The category isn't niche anymore , it's setting the pace for the rest of tokenized assets.

openai briefly leaked model names in the codex picker, including gpt-5.5, spud, and oai-2.1. the screenshots vanished fast, but it is a pretty direct sign that the next model line is already being wired into codex surfaces.

The evidence gap is stark. MOTS-c has published mitochondrial and metabolic research plus early human pilot data. BPC-157 has no FDA-registered human efficacy trials and relies mostly on older animal wound-healing studies. Both are being judged under the same 503A standard anyway. If a peptide stays off the list, bulk compounding is blocked. The committee vote does not finalize policy, but it sets direction before FDA decisions and follow-on moves by state boards and payers. July 23-24 is where access to these four peptides starts getting decided.

Pod state is replicated across members with Raft consensus, including membership, API keys, and a shared cloud budget. Hyperspace is pitching it at teams already spending $500 to $2,000 a month on API calls, with the idea that daily coding, research, and drafting work can run on hardware they already own. Cloud still stays in the loop for jobs that need a frontier model, but local prompts are supposed to cost only the machines' electricity.

The claimed gains are not small: up to 58% faster time-to-first-token and 45% faster time-per-output-token versus W4A16. The tricky part was quality, FP8 scale casting was clipping, so they switched to per-channel quantized scales with a 1/8 rescale, recovering over 99.5% of W4A16 accuracy on Command A and Cohere MoE. This puts W4A8 directly into production inference pipelines.

Reports say Cursor gets massive compute access and a path to train its own model, while SpaceX/xAI gets a dominant coding app with real distribution and ~$2B ARR. The play is owning both layers at once. If the model catches up, they buy the stack. If not, they still capture demand, data, and product learnings.

At the same time, writes moved the other way. Standard API posts went from $0.01 to $0.015, URL posts jumped to $0.20 except summoned replies at $0.01, and follow, like, and quote actions were pulled from self-serve tiers. Net effect: cheap inward data loops, pricier and tighter outbound actions, with official agent tooling and tutorials arriving alongside it.

At the same time, RM-055 showed up as a different mechanism entirely: a mutant-targeted catalytic RAS(ON) inhibitor. Preclinical work showed deep, durable antitumor activity in pancreatic models, positioned as a way to get around resistance rather than just improve on the first wave.

Administration allies are framing it as accelerated clinical access, not retail prescribing. The goal is to get supervised treatments into practice sooner if trials line up. Supporters point to a Stanford veteran cohort reporting large symptom drops one month after treatment: PTSD down 88%, depression 87%, anxiety 81%. That data is being used to argue urgency. The order leans on regulators to expand trials and define how these therapies are used. The immediate mechanism is bureaucratic pressure. Faster funding decisions, more trial support, and shorter approval timelines aimed at turning small studies into sanctioned treatment options.

The mechanism is forced proximity. These molecules create protein pairings the cell did not evolve for, then route them to degradation. One case targets nuclear pore components. Another targets histone deacetylases through KBTBD4. Another goes after BRD9 through DCAF16. Liau also pointed to deep mutational scanning as a way to systematically find these interactions. It uses mutation-driven binding as a discovery engine instead of one-off hits.

Eight patients mounted strong vaccine-linked T‑cell responses, and nearly all of them were still alive at about the six-year follow-up. That's against a backdrop of roughly 13% five-year survival overall. It's a tiny phase 1 cohort, but the signal is durability in a cancer where durable responses are rare.

The problem was mix and pressure: lower-margin AI work, a 31.5% non-GAAP operating margin guide, and renewals dipping to 97%. Add a 0.75 pt margin hit from the Armis deal and a 0.75% Q1 growth headwind from delayed Middle East deals, and traders hit the whole group. IGV exposure helped drag Oracle, Salesforce, Adobe and others down with it.

The pitch is volume at creator-team scale. Demos and posts claim 500+ videos per day, nine formats from one link, and costs framed under $1 per campaign or $39/month versus a $50k+ creative team. Hermes pulls hooks from Meta and TikTok ad libraries, then generates face-locked avatars and usable product footage for web and app campaigns.

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.

The key step is that Smart Mesh avoids the usual messy geometry and edge flow problems, so there's less cleanup before import. That's how it jumps from prompt to playable space with AI behavior and collisions already working.

That leaves the headline concern looking a lot less straightforward. Muscle biopsies in the newer data pointed to only a 10 percent loss in actual functional tissue, while a separate meta-analysis found GLP-1 users losing muscle at the same rate as patients relying only on caloric restriction.

Instead of retraining a whole LLM, BAR updates pieces separately and stitches them back together. It takes the modular setup from FlexOlmo, which handled pretraining, and pushes it into post-training, where teams usually eat full retrain costs. The pitch is continuous improvement and setups where data can't be pooled. Allen AI dropped models, a blog, and a paper alongside the release.

The idea is opposing effects: tirzepatide improves glucose but nudges HR up, while CJC-1295 boosts GH/IGF-1 but worsens insulin resistance. He said he'll compare a full post‑CJC week to baseline and abort or switch to no‑DAC CJC-1295 plus ipamorelin if the sleep disruption persists.

Arm's estimate: a 1GW AI datacenter jumps from about 30 million CPU cores to roughly 120 million for agent-heavy workloads, with agents generating more than 15x the tokens of human users. The pitch behind its new AGI CPU is simple: pack far more CPU throughput into the same power envelope as operators try to keep up.

The shift is variance. The AI runs cluster tightly with no extreme outputs, while the human sample spreads wide. The paper frames that consistency as useful for scaling structured empirical work: you get middle-of-the-pack results, but you get them reliably.

The bigger pattern they point to is Bedrock deprecations not lining up with real availability. Their groups haven’t seen a clean, permanent loss of an Anthropic model since Claude Instant 1.2, and even after Sonnet 3.5 and 3.6 were removed on April 1, they had previously found older models still reachable in certain AWS regions.

First human doses are getting moved offshore, not just the lab work. Australia is pulling in early biotech trials because startups can clear ethics and regulators faster and start dosing sooner, with less cost drag than the US. That's a direct response to Eroom's law: AI helps discovery, but safety still needs real humans. One comparison puts Phase I at ~$300M in the US vs ~$20M in India, with Australia emerging as a middle path plus lighter trial setup.

More than 30% of patients prescribed specialty drugs never even start, not because of the medicine but because the support system breaks before treatment begins, per Oak HC/FT. Courier Health is going after that gap with software that ties together manufacturers, pharmacies, and patient support workflows, and just raised a $50M Series B led by Oak HC/FT to scale it.

Anthropic is moving to lock down its own compute in Europe instead of leaning on cloud access. A new six-figure business development role is focused specifically on negotiating data center and capacity deals in-region, with chatter pointing to leasing efforts in places like London. That shifts Europe from a future market into an active front in the AI compute bottleneck, with Anthropic directly procuring infrastructure.

PerfectSquashBench is meant to capture that lock-in: image models anchor on their own prior outputs more than text models do. The comparison is simple: keep iterating and you get near-copies, reset context and it breaks. Mollick's joke is that this is how models end up "squashmaxxed," optimized for one narrow visual groove instead of actually improving across iterations.

An LLM handled an obscure AML/KYC literature search and actually turned up internal prosecution guidance without much hand-holding. In a test, Opus 4.7 found multiple relevant government docs on its own and lined up with the idea that these rules act as process-crime tripwires for prosecutors. The only failure was mechanical: it couldn't fetch one known PDF, which had to be supplied manually. That shifts the bottleneck from discovery to retrieval.

Frontier AI labs are actively buying their way into biotech data instead of relying on public benchmarks, cutting deals with startups to access proprietary wet-lab and experimental datasets. The shift is happening at scale, basically every startup with a real high-throughput platform has been approached for licensing. The edge is moving to whoever locks up the most closed, continuously generated lab data before bio models commoditize.