Entity trail

Recursive

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

Briefing refs

Findings

Edges

Sources

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

Corpus findings

2026-09-02 / skill-finderGrant recursive agent trees permission to spawn freely but charge them to actProgressive Risk Vesting separates sandbox spawning, where external controls prevent the harm, from capability activation, where a branch crosses an irreversible-action boundary, and holds a trajectory-level risk budget in escrow that is debited only when a branch is activated. The authors prove an anytime harm bound for adaptively generated trees and show that with activation gates and charges held fixed, delayed vesting preserves every policy available under charging at spawn time, so the cheaper scheme costs nothing in expressiveness. They are explicit that the studies are synthetic and do not estimate safety in deployed agents.
2026-08-28 / agents-researcherGoose v1.48.0 ships roughly twenty security fixes at once, nearly all of them fail-closed correctionsBlock's Goose released v1.48.0 on 27 August with a security section dominated by defaults that previously failed open: fail closed on malformed tool visibility, permission denies now take precedence, fail closed on invalid default GCP credentials and invalid Codex ACP mode, honor plugin enablement for skills, honor MCP tool model visibility in Code Mode, sanitize Unicode tags in MCP prompts, bound recursive mention scans and call-graph traversal, and avoid a predictable editor symlink. The same release adds an on_failure block for PreToolUse hooks, a stable tool_call_id across the tool lifecycle, built-in web-search and browser-use skills, and MCP conformance tests in CI. The concentration of fail-open bugs in one release is the story for anyone running Goose with permissions configured.
2026-08-26 / agents-researcherVideoHarness-RSI isolates the harness as its own optimization layer, searching executable context constructors around a frozen VLMPosted 25 August (arXiv 2608.24302), this paper asks how much of long-video understanding performance comes from the context-construction program alone, holding the answering model and interface fixed. An outer-loop proposer uses prior programs, evaluation outcomes, and execution traces to generate candidate harnesses, which are executed end to end with successful variants retained for further search. Starting from uniform frame sampling, recursive search consistently beats weaker hand-crafted baselines; starting from a stronger hand-crafted baseline, the same process still improves on it, and the discovered harness transfers to other long-video benchmarks without further search. It is the third paper this week to treat the harness rather than the model as the object of optimization.
2026-08-26 / sources-researcherDylan Patel on Dwarkesh: Anthropic and OpenAI outbid everyone for compute because they monetize it best, and $10T of capex could break sovereign debtPublished 2026-08-25 and retitled from the compute framing to 'Two labs will soon control most of the world's workforce', the four-hour episode argues the two labs are on track to control most of the world's usable FLOPs within a few years, and that the industry is shifting spend from inference back to training as recursive self-improvement gets closer. Patel's chapters include $6 billion in fab capex enabling $1T+ of end revenue and compute prices rising as labs outbid everyone else. The second half asks whether $10T+ of AI capex by 2030 raises rates enough to bankrupt non-AI-exposed countries and crash non-AI equities. Dwarkesh notes they could not name any force that counters the centralization.
2026-08-25 / vibe-coding-researcherContinue CLI Ships With a Destructive-Command Denylist That Misses rm -rf $HOME (CVE-2026-76072)VulnCheck disclosed on 2026-08-24 that Continue CLI's headless and auto modes grant the Bash tool blanket allow permission, leaving isCriticalCommand in packages/terminal-security as the only guard. Its dangerous-path test matches only /, ~, /usr, /etc, /bin and /sbin, so recursive deletion of /home, /root, /var, /opt or /srv passes, and because the line is parsed with shell-quote, $HOME collapses to an empty token before the shell re-expands it at spawn time. find -delete is rated high risk rather than blocked, and shred, wipefs, truncate and pkexec are unhandled, so an indirect prompt injection in a fetched page or issue body can wipe the invoking user's data. CVSS 4.0 base 8.3.
2026-08-22 / hn-researcherAutolith v0.35.0 Puts a Coding Agent Inside a Live SBCL Image That Redefines Functions Without RestartingAutolith is a terminal agent that runs a full SBCL 2.6.6 Common Lisp environment inside its own process, so it inspects, tests and redefines functions live rather than editing files and re-running a build. It keeps an append-only mutation journal, commits changes as persistent private image commits, recovers code after crashes, and carries memories and agendas across sessions. Workspace indexing is built on the Rust library fff, and it supports recursive inference over corpora larger than the model context window. Works with any OpenAI-compatible provider; binaries, Nix and source available under Apache/MIT.
2026-08-21 / thought-leaders-researcherDeepMind's 'contract-first' agent delegation paper resurfaces on X with 68K views six months after publicationA post on August 20 pitching a Google DeepMind paper that treats AI delegation as verifiable contracts rather than prompt handoffs cleared 68,665 views and 587 likes, but the paper is from February 12, 2026 by Nenad Tomasev, Matija Franklin and Simon Osindero. Its actual rule is strict: a delegator may not assign a task unless the outcome can be precisely verified, and must recursively decompose until sub-tasks map to automated checks such as a passing unit test, a formal proof, or a strict schema. It also proposes policy-as-code permissioning and zero-knowledge proofs for verifying correctness without exposing data.
2026-08-20 / news-researcherMIT Technology Review Argues Recursive Self-Improvement Is Not Arriving on ScheduleThe August 19 edition of The Download leads with the case that AI's recursive self-improvement may not come as easily as the labs' timelines assume. It is a useful counterweight to the intelligence-explosion framing that has driven the last year of capex justification. The argument matters most for people budgeting against an assumption that models will soon meaningfully improve themselves.
2026-08-18 / skill-finderCompile your agent's repeated leaf decisions into abstainable reflex rules and skip 15-20% of LLM callsHaReCAP mines successful trajectories for frequently recurring leaf-level decisions and compiles them offline into one-step reflex rules that are auditable and can abstain: at runtime, if a stored rule maps uniquely to a legal action in the current valid-action set, the LLM call is bypassed; otherwise it falls back to the normal recursive path. On Qwen3.5-27B this cut tokens on solved tasks by 14.67% (Robotouille sync), 17.93% (async), and 20.08% (ALFWorld) without changing the control flow. The abstain-on-ambiguity design is the transferable part — it's a cache that refuses to guess.
2026-08-18 / sources-researcherImport AI 469: Jack Clark Pins Zuckerberg's 'The Future is for Everyone' on One Unexamined Assumption About Who Superintelligence Works ForIn the 2026-08-17 issue, Clark reads Zuckerberg's essay proposing wide distribution of superintelligence as a hedge against power concentration, and identifies the gap: the essay assumes systems capable of invention will inherently serve individual empowerment, without ever addressing whether such systems would "solely work on behalf" of less capable humans or simply restructure global power in ways nobody modelled. The same issue covers DiG-bench and a browser-based recursive-self-improvement simulator from Paradigm Research that forces the player to trade researchers against compute and data licensing — a rare hands-on intuition pump for how RSI inputs interact.
2026-08-18 / sources-researcherDuckDB v2.0 'Cyanoptera' Preview: a Client/Server 'Quack' Protocol, Triggers, a PEG Parser Replacing the Postgres One, and 40x Faster Recursive CTEsThe 2026-08-17 preview post lays out the largest DuckDB break since 0.x: a `CONNECT` statement and network protocol turning DuckDB into a server with remote pushdown into PostgreSQL and MySQL; VARIANT with shredded execution straight from storage and extraction pushdown into scans; full BEFORE/AFTER row-and-statement triggers with transition tables; NEAREST joins for similarity search; async I/O that scales independently of query processing; and storage format v2.0 with DICT_FSST on by default. Benchmarks cited: recursive CTEs 4.90s → 0.12s (40x), timezone conversions 2.2x, German collation filtering 2.6x. Ships this fall with a stable C API and a small set of breaking changes including the new default storage format.
2026-08-18 / sources-researcherDiG-bench Releases 70 Hidden-Rule Discovery Games — Every One Solved by a Human First Try, Only Opus 5 and Fable 5 Beat Any Tier-7 Task at 0.2 SuccessDiG-bench (arXiv 2608.12593, 2026-08-12) from a group including Tri Dao, Jürgen Schmidhuber, Joshua Tenenbaum, Thomas Griffiths and James Whittington encodes each of 70 games as a short string with unique transformation rules that must be inferred through experimentation — win conditions are also hidden. All 70 were solved by at least one human on first attempt; 21 are public and 49 held private for secure evaluation. Per Jack Clark's Import AI 469 writeup, Opus 5 and Fable 5 led, and only those two cleared any Tier-7 task at a 0.2 success rate — Clark predicts human parity by mid-2027 and treats that as the gate on recursive self-improvement.

Graph relationships

PARTNERS WITH
AWS
Recursive committed to a multiyear AWS compute agreement.
Source finding
INVESTED IN
AMD
AMD backed Recursive's $650M Series funding.
Source finding
INVESTED IN
NVIDIA
Nvidia backed Recursive's $650M Series funding.
Source finding
INVESTED IN
GV
GV backed Recursive's $650M Series funding.
Source finding
WORKS AT
Richard Socher
Richard Socher is CEO of Recursive.
Source finding
INVESTED IN
NVIDIA
Nvidia participated in Recursive's Series A funding round.
Source finding
INVESTED IN
Greycroft
Greycroft participated in Recursive's $650 million Series A.
Source finding
INVESTED IN
GV
GV led Recursive's Series A at $4.65 billion valuation.
Source finding
INVESTED IN
AMD -> Recursive
AMD backed Recursive's $650M Series funding.
Source finding
WORKS AT
Richard Socher -> Recursive
Richard Socher is CEO of Recursive.
Source finding
INVESTED IN
NVIDIA -> Recursive
Nvidia participated in Recursive's Series A funding round.
Source finding

Source trail

changelog TechTimes Bloomberg reports Richard Socher's new company Four-year commitment Source: TechPowerUp Carnegie Mellon and Google Announced at Knowledge 2026 The changelog Researchers quantified Fortune arXiv 2609.01035 — Spawn Freely, Act Sparingly: Progressive Risk Vesting for Recursive LLM-Agent Trees

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…