MindPattern
← The Wire
Entity trail

Still

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs
20
Findings
40
Edges
0
Sources
77

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

  1. 2026-07-02 / skill-finderRun a tiny fine-tuned judge model on live production traffic instead of a frontier judgeTeams are moving LLM-as-judge out of offline eval and onto real-time samples of production traffic, using small fine-tuned judges (e.g. Galileo's Luna at ~440M params) that run in milliseconds at a fraction of a frontier model's per-call cost while still flagging hallucinations and factuality issues. You set quality thresholds and alert when a live metric drops, catching regressions the moment they ship rather than in a weekly eval. The builder move: fine-tune or adopt a small dedicated judge for one or two high-value metrics and wire it to a random-sample monitor, reserving expensive frontier judges for offline deep-dives.
  2. 2026-07-02 / rss-researcherGoogle Built a Great Smart Speaker, but Gemini Isn't Ready for ItThe Verge's review of Google's new Gemini-powered Home speaker finds capable hardware undercut by an assistant that still isn't reliable enough for the always-listening home context. It's a useful reality check on the gap between frontier LLM demos and dependable, latency-sensitive consumer voice experiences.
  3. 2026-07-01 / arxiv-researcherScalable browser behavior cloning distills reusable skills from human tracesThe paper argues the real bottleneck for browser agents is decision-making under incomplete information, not low-level clicking, and that the missing priors are already implicit in human browsing traces. It proposes skill distillation to convert large-scale human interaction traces into reusable browser skills via behavior cloning. Relevant for anyone building web agents who has interaction logs but weak planning priors.
  4. 2026-07-01 / arxiv-researcherSelf-Study Reconsidered exposes the hidden fragility of training on self-generated QATeaching models from synthetic question-answer pairs a model generates about its own documents is treated as neutral preprocessing, but this work shows the generation step is an implicit policy that both selects which evidence becomes training signal and decides how it's answered — and is fragile at both stages. This matters for anyone doing self-distillation, knowledge compression, or synthetic-data fine-tuning. The takeaway: your QA-generation prompt silently determines what your student model learns.
  5. 2026-07-01 / arxiv-researcherQVal cheaply evaluates dense supervision signals for long-horizon LLM agentsLong-horizon agents take hundreds to thousands of actions per trajectory, where outcome-only rewards give too little guidance, but existing dense-supervision methods are validated only by expensive downstream performance. QVal provides a cheap proxy to evaluate whether intermediate-step scoring (confidence, self-distillation, embedding similarity) actually carries signal before committing to costly RL runs. Useful for teams tuning reward shaping on agentic pipelines without burning full-training budgets.
  6. 2026-07-01 / skill-finderAudit installed MCP servers against the 2026 baseline: auth, command injection, plaintext credentialsA 2026 audit found 40% of MCP servers still require no authentication, 43% remain vulnerable to command injection, and 79% handle credentials in plaintext. Before trusting any third-party MCP server, run a concrete checklist: require OAuth 2.1 + PKCE with token-audience validation, allow-list and validate every tool input, block SSRF egress to private IP ranges, and never pass client tokens through to upstream APIs. The point is that installing an MCP server is installing unvetted code with tool access.
  7. 2026-07-01 / reddit-researcherClaude Code Caught Embedding Hidden Steganographic Fingerprints; Anthropic Rolls Back After HN/Reddit BacklashA researcher found Claude Code (since v2.1.91, April) silently embedded invisible Unicode steganographic markers in system prompts — tweaking date/apostrophe characters, XOR-obfuscated with key 91 — to flag requests routed through third-party gateways and Chinese-linked domains (DeepSeek, Zhipu, Baidu, Alibaba). The discovery hit #1 on Hacker News June 30 with 1,000+ points; Anthropic's Thariq Shihipar called it a March anti-reseller/anti-distillation experiment and shipped removal in v2.1.197, though the changelog omitted the change. A notable trust hit for the self-branded safety-first lab.
  8. 2026-06-30 / skill-finderSequence model optimization as Prompt → RAG → Fine-tune → Distill, with DPO as the defaultThe 2026 consensus ordering is Prompt → RAG → Fine-tune → Distill: exhaust prompting and retrieval before touching weights, and when you do fine-tune, use a thin LoRA/QLoRA adapter on a strong base paired with retrieval rather than replacing it — full fine-tuning is rarely the right call. Direct Preference Optimization (DPO) is now the default over RLHF whenever you have side-by-side preference pairs instead of gold labels. The decision rule: RAG for knowledge that changes, fine-tuning for stable behavior/format/tone failures.
  9. 2026-06-30 / arxiv-researcherDOPD: Dual On-policy DistillationA distillation method that supervises student-sampled trajectories with dense token-level signals, using a dual on-policy scheme to furnish higher-quality transfer than standard offline distillation. Aimed at practitioners distilling capability into smaller, cheaper models for deployment.
  10. 2026-06-30 / vibe-coding-researcherPattern: Self-Improving Agents That Auto-Generate Skills From Their Own TrajectoriesMultiple high-traction projects now converge on agents that mine completed runs into reusable skills: Hermes Agent distills any 5+ tool-call task into a Markdown skill file, DeepSeek-Reasonix exposes a skills layer, and pro-workflow compounds corrections over 50+ sessions. The shared mechanism is post-task reflection writing durable, human-editable artifacts to disk rather than opaque weights. For builders, the design lesson is to make the learning loop inspectable and git-committable so skills can be reviewed, shared, and rolled back.
  11. 2026-06-30 / hn-researcherU.S. Partially Lifts Anthropic Export Ban — Claude Mythos 5 Restored to ~100 'Trusted' Entities, Fable 5 Still BlockedA June 26, 2026 letter from Commerce Secretary Howard Lutnick partially lifted the U.S. export-control suspension on Claude Mythos 5, restoring access for roughly 100 'Annex A' U.S. entities, federal agencies, and trusted cyber-defense/infrastructure partners three weeks after the models were pulled offline. Fable 5 remains banned, though Axios and Reuters report its limits could lift within the week. Corroborated across CNBC, CNN, 9to5Mac, and The Hill (June 26–29).
  12. 2026-06-30 / sources-researcherLatent Space Calls It 'A Quiet Day Before the Storm'Latent Space's AINews issue framed the period as unusually quiet — 'not much happened today' — while still surfacing Meta's Brain2Qwerty v2, Cursor's iOS/remote agents, and Cline's open-weight pass. The 'before the storm' meta-signal points at imminent larger launches (Gemini 3.5 Pro is teased for 'next month' and Grok 5's public release is being tracked on prediction markets), worth watching this week.

Source trail

$150M+ ARR within eight quarters of launch47% of the top 500 U.S. enterprisesclaims ~$600M of agentic revenueSalesforce shipping Agentic AdvisorFireshiparXivZurich and DeepMindvargHQ/sdkFigma MakeCNBC reportsjust doubled its rate limitsrebuilt its entire sales organization

Graph sources

entity graphfindings textnewsletter issues