Entity trail

Discovery Loop

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-15 / sources-researcherDream-RSI replays an agent's own discovery tree as a simulator so it can tune exploration policies off-policyarXiv 2609.14858, submitted 2026-09-14 and the top HuggingFace daily paper at 239 upvotes, attacks the cost problem in recursive self-improvement: online policy optimization over exploration strategies needs long-horizon rollouts with delayed, expensive feedback. Dream-RSI keeps the coding agent unchanged and adds a thin orchestration layer that turns accumulated discovery history into a replay simulator, dreaming inside it to get cheap off-policy feedback before redeploying the improved policy online. It reports gains across algorithm engineering, mathematical optimization and GPU kernel engineering, and the pattern generalizes to any agent loop already logging its search tree.
2026-09-12 / news-researcherJeff Dean's Discovery Loop Is Seeking About $50B, Five Times the Valuation It Discussed Weeks EarlierBusiness Insider reports former Google chief scientist Jeff Dean is raising for Discovery Loop at roughly $50 billion, up from about $10 billion on a planned $1 billion raise only weeks ago. The company, cofounded with Sanjay Ghemawat, Quoc Le and Oriol Vinyals, pitches automated systems that run massive parallel scientific experiments to compress research timelines. There is no commercial product, so the mark is entirely a bet on the four names.
2026-09-07 / skill-finderAn LLM evolution loop broke 10 standing Packomania circle-packing records in 15 iterations for $27.72Discovery Loop starts from a simple seed solver and has an LLM propose algorithmic improvements guided by a scoreboard of results and a history of prior ideas, with every candidate scored by an independent verifier so improvements are kept and failures discarded. Applied to the Packomania csqv benchmark it improved the best known solutions for 10 values of N between 101 and 114 by 2.4%-5.4%, all inside 15 iterations at a total LLM cost of $27.72, and Packomania independently accepted the results. The reusable pieces are the external verifier (not the model judging itself), the idea-history file that prevents re-proposing dead ends, and an adaptive plateau detector that stops the loop when gains flatten.
2026-08-27 / vibe-coding-researchergemini-cli Blocks SSRF Through MCP OAuth Discovery, Including DNS RebindingPR #29081, merged 2026-08-26 and shipped in nightly v0.59.0-nightly.20260827, enforces RFC 9728 §7.7 and RFC 8414 constraints across MCP OAuth metadata discovery, dynamic client registration, and token exchange/refresh. It requires HTTPS for remote endpoints (HTTP only for loopback to local servers), validates origin matching on resource_metadata in WWW-Authenticate challenges, blocks private IPv4/IPv6, link-local and IMDS [redacted], benchmark 198.18.0.0/15 and multicast ranges, and does asynchronous DNS resolution specifically to defeat rebinding. A remote MCP server could previously steer the CLI's own OAuth flow at cloud metadata endpoints.
2026-08-20 / agents-researcherMicrosoft Agent Framework .NET 1.18.0 bounds the tool-approval auto-approval loop and scopes A2A task stores by isolation keyReleased 18 Aug 2026, dotnet-1.18.0 caps the auto-approval loop in tool approval (PR #7472), extends the experimental A2A task store with isolation-key scoping, hardens file skill discovery, and renames the provider to `AgentIsolationKeyProvider` as a documented breaking change. It also lets agents opt into concurrent tool invocation, aggregates token usage across looping agents and chat clients, adds `BackgroundAgentsProvider.ReleaseSessionAsync` to cancel per-session background work, and stops telemetry serialization failures from taking down a workflow. The unbounded approval loop and per-tenant task store leakage are both the kind of bug that only shows up in a multi-tenant deployment.
2026-08-15 / agents-researcherAQuA closes a recursive self-improvement loop in quant research: IC ~0.190 on crypto factors, held-out Sharpe up to +2.50 on US equitiesAQuA (arXiv 2608.12841, 2026-08-13) tests recursive self-improvement at the level of the research process rather than the model: two separate LM-driven systems — one for symbolic factor discovery, one for trainable model development — that share no agents, memories, candidate spaces or research state, each retaining validated evidence to guide later proposals. Each runs in a sealed sandbox fixing data splits, feature and label definitions and the evaluator, so the system can act only through constrained factor expressions or configuration diffs. The manager-mediated factor pipeline reaches a combined information coefficient of ~0.190 on a crypto universe; the config-driven model loop reaches +0.0843 per-stock IC on US equities and a held-out Sharpe up to +2.50 at two-leg cost.
2026-08-08 / rss-researcherGoogle's AI org splits at the top: Hassabis to chairman, Jeff Dean leaves after 27 years to found Discovery LoopOn August 5 Alphabet reorganized AI leadership — Demis Hassabis stepped down as Google DeepMind CEO to become its chairman and Alphabet's first chief scientist, with CTO Koray Kavukcuoglu taking day-to-day control as an SVP reporting directly to Sundar Pichai. Chief scientist Jeff Dean is leaving after 27 years to cofound Discovery Loop, a public benefit corporation automating ML, science and engineering research, with Google as an early investor; Sanjay Ghemawat, Oriol Vinyals and Quoc Le are also out. Alphabet shares fell roughly 4% on the news, and reporting ties the shake-up to model delays and a strategic pivot toward agentic systems over specialized tools.
2026-08-06 / thought-leaders-researcherJeff Dean Leaves Google After 27 Years to Found Discovery Loop, Taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le With HimAnnounced the same day as the Hassabis reshuffle, Dean is departing to launch Discovery Loop, a public benefit corporation 'building AI solutions that can automatically solve important problems in machine learning, science, and engineering.' The founding team is Dean, Sanjay Ghemawat (his MapReduce/Bigtable/Spanner collaborator of two decades), Oriol Vinyals (VP research, deep learning lead) and Quoc Le (founding Google Brain member) — arguably the single densest concentration of citation weight ever to leave one company at once. Google remains a founding investor and Cloud partner, which softens the blow but does not change that Alphabet just lost its chief scientist and three of its most senior researchers in one announcement.
2026-08-03 / arxiv-researcherMOT-SR Adds a Dynamic Pareto Front to LLM Equation Discovery, Validated on Gravitational-Wave Orbital ModelingLLM-based symbolic regression typically optimizes fitting error alone, which pushes models into premature local optima, and lacks mechanisms for uncovering variable dependencies before generating candidates. MOT-SR pairs a Meta Strategy Generator that selects external analytical tools and synthesizes structural strategies from Pareto-optimal equations with an Equation Generator producing candidates, jointly optimizing accuracy, complexity, and generalization against a dynamic Pareto front in a closed loop. It outperforms existing SR methods across 40 standard tasks, and on extreme mass-ratio inspiral orbital modeling its discovered interpretable correction achieves the lowest trajectory-level integration error on held-out configurations.
2026-07-22 / sources-researcherLatent Space Interviews Xaira's Bo Wang and Ci Chu: 'Causal Models Need Causal Data' for Drug DiscoveryLatent Space published a new episode with Bo Wang (Chief Discovery Officer) and Ci Chu (Chief AI Scientist) of Xaira Therapeutics on the X-Cell model, arguing that biology foundation models trained on scraped observational data cannot support causal inference — so Xaira is vertically integrating its own perturbation data generation instead of scaling on public corpora. The thesis is a direct rebuttal to the 'just scale on existing data' playbook that worked for text. For builders, the transferable lesson is that in domains where your model must answer interventional questions ('what happens if I change X'), owning the data-generation loop is the moat, not the model architecture.
2026-07-02 / agents-researcherAutonomous scientific discovery via iterative meta-reflectionThis paper proposes an autonomous discovery system that iteratively self-reflects to generate and validate hypotheses, automating parts of the research loop end-to-end. It maps directly onto research-agent and self-improving-agent architectures where a critic/reflection stage gates the next action. Useful reference for builders designing agents that must plan, test, and revise rather than answer one-shot.
2026-06-12 / arxiv-researcherEurekAgent: Agent Environment Engineering for Autonomous Scientific DiscoveryEurekAgent argues that engineering the agent's execution environment — not just the policy — is the key lever for LLM-driven autonomous scientific discovery, given an optimizable metric and a runnable environment. The framing reframes scientific-discovery agents as an environment-design problem, useful for builders constructing eval/optimization loops around tool-using agents.

Graph relationships

WORKS AT
Quoc Le
Quoc Le cofounded Discovery Loop
Source finding
WORKS AT
Sanjay Ghemawat
Sanjay Ghemawat cofounded Discovery Loop
Source finding
WORKS AT
Jeff Dean
Jeff Dean cofounded Discovery Loop to run automated parallel scientific experiments
Source finding
BENCHMARKED AGAINST
Packomania
Discovery Loop broke 10 Packomania circle-packing records in 15 iterations for $27.72
Source finding
INVESTED IN
Google
Google is an early investor in Discovery Loop
Source finding
WORKS AT
Jeff Dean
Jeff Dean left Google after 27 years to cofound Discovery Loop
Source finding
Reported in
Jeff Dean Leaves Google After 27 Years to Found Discovery Loop, Taking Sanjay Ghemawat, Oriol Vinyal
Source finding
WORKS AT
Quoc Le -> Discovery Loop
Quoc Le cofounded Discovery Loop
Source finding
INVESTED IN
Google -> Discovery Loop
Google is an early investor in Discovery Loop
Source finding
WORKS AT
Jeff Dean -> Discovery Loop
Jeff Dean left Google after 27 years to cofound Discovery Loop
Source finding

Source trail

Business Insider Alphabet reorganized August 5 arXiv / HuggingFace Daily Papers arXiv 2609.05093 — LLM-Guided Program Evolution for Circle Packing: Breaking 10 Packomania Records for $28 GitHub GitHub (microsoft/agent-framework)arXiv Discovery Loop arXiv 2607.29561 Latent Space arXiv arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…