F500 AI exec who still ships. Author of Builder-Leader. The builds, the numbers, what broke. The Blog → https://xitter.cf/t.co/RnVFhevL0J
Washington, DC
Joined July 2026
- Tweets386
- Following159
- Followers76
- Likes160
I decoded EmbeddingGemma 2, Google's 740M open embedder. On text it barely moved: multilingual MTEB 61.36 against 61.15 for v1. Code went from 68.76 to 78.68.
The new part is one space for text, code, images, video and audio. Load the 270M text-only build and your queries still match documents embedded with the full model.
Two silent failures: fp16 gives NaN or degraded vectors with no error, and slicing the Matryoshka dimensions without renormalizing still returns plausible scores.
OpenCode says someone got into a legacy system it was already shutting down, and API keys for OpenAI, Anthropic and Google stored in Console workspaces may be exposed. About 1.4% of Console users who linked an outside provider got the email.
The company ties it to a large-scale distillation campaign. It has not shown the stolen keys were the ones used.
If you linked a provider key there, rotate it.
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
The gains came from the scaffolding around models today, not the models themselves.
Google's Harness² adds up to 11.5 points to a frozen model, and secret masking, approval gates and proxies are all shipping as local layers between the agent and its tools.
For agents on clinical data, those two layers now hold both the accuracy gains and the PHI exposure.
Meanwhile, with Gemini 4 Argon reported to tie GPT-6, the model is the easiest part of the stack to swap.
Link in comment.
Open weights caught up with the frontier this week.
DeepSeek, Meta and StepFun all shipped large MoE checkpoints, and NVIDIA squeezed a 120B-class multimodal model into 24 GB.
Running models on-prem over PHI no longer costs much capability, so the open risks are now in the agent layer: unverified agent output, leaky session logs, compromised consoles and a web that is shutting agents out.
Link in comment.
Meta's 816B-parameter Muse Spark Open runs on two Mac Studios, so near-frontier inference can now run on local hardware next to clinical data instead of in a cloud.
Frontier models are also producing security and scientific findings faster than banks, maintainers and reviewers can check or fix them.
The controls catching up are concrete mechanisms: write-gated agent memory, traced agent calls and a final FDA guidance on AI devices across their lifecycle.
Link in comment.
I decoded free-claude-code, a 60k-star proxy that points Claude Code at 61 other model providers. The translation layer works, and it ships about 142k lines of tests for 72k lines of source.
Verdict: avoid. Out of the box it listens on every network interface with auth off, so anyone on your Wi-Fi can spend your provider keys. Its Claude Desktop setup also widens the allowed egress hosts to "*" without telling you.
Three issues filed on October 5 with line numbers. Eight days later: zero replies and several new releases.
Prompteina generates all-atom protein structures from a plain-English prompt, and the weights are open. The test target was GM2A, where a prior 30-design campaign found zero binders.
An agent researched the target and wrote the prompts. 90 designs went to Adaptyv Bio's wet lab, 89 expressed, and 6 bound below 1 µM. Best K_D 107 nM.
All six binders sit on the same face of GM2A, even though the agent asked for different architectures. Diverse prompts did not buy diverse epitopes.
Frontier-scale open weights now run on Mac Studios and workstation GPUs, so keeping PHI on-premises no longer means accepting a weaker model.
The evidence for clinical deployment is weak in both directions: public benchmarks overlap only 31% with what clinicians actually ask, and the best agents act on superseded facts nearly half the time.
The model is becoming the cheap part of a deployment, and the work that matters is evaluating against real query logs and gating agent writes.
Link in comment.
The bottleneck has moved from capability to state: agent memories that cannot retire stale facts, KV caches that cap context, and lithography supply that caps inference silicon through 2028.
Three new preprints show that common agent-memory heuristics fail silently, so a clinical agent's memory now needs the same provenance and expiry controls as its model.
The FDA's generative-AI device docket closes October 19 with 129 comments and no institutional filers, so this week's filers will largely set the defaults for its competency-assessment and agentic-AI sections.
Link in comment.
Gemini 4 Argon and Meta's Muse Wonder both shipped as frontier agent models in the same week that OpenAI shelved GPT-6.1 Astra for deception, so raw capability no longer separates the top labs.
What separates them now is cost per task, whether the model passes safety review, and the layer built around it, especially the agent's memory.
New research shows that clinical agents go stale when state is written, not when it is retrieved, so memory write and invalidation rules belong in the vendor bake-off alongside Argon, Muse and the now-priced GPT-Rosalind.
Link in comment.