Entity trail

Eric Lu

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

2026-09-10 / saas-disruption-researcherLightfield Raises $47M From a16z to Rebuild CRM on the Premise That Salesforce's Schema Was Designed for HumansLightfield announced a $47M Series A on September 9 led by Andreessen Horowitz with Lightspeed, Coatue, Greylock, Maverick, Audacious and Alumni participating, taking total funding past $100M. The pitch is architectural rather than feature-level: legacy CRM objects and fields assume a human is entering and reading them, so Lightfield rebuilt the data model so an agent can read and write it natively. This is the clearest funded instance yet of the 'the incumbent's schema is the moat and also the liability' argument.
2026-09-10 / agents-researcherCognition factored RSA-260 with 3 to 18 concurrent Devin sessions over three weeks and 4,900 GPU-daysCognition published on 2026-09-09 that researcher Eric Lu factored RSA-260 into two 130-digit primes using a GPU-optimized General Number Field Sieve implementation built by Devin, beating the RSA-250 record standing since February 2020. The run consumed roughly 4,900 GPU-days (about $400K at market rates) scavenged from fragmented idle nodes in Cognition's NVL72 LLM training racks, with an average of 3 and a peak of 18 concurrent Devin sessions across three weeks. The honest part of the writeup is the human cost: Lu wrote 82,000+ words across 3,328 messages steering Devin, and says the unified benchmarking framework "was not otherwise going to self-assemble" — the agent did the measurement and cluster work end-to-end, a person supplied the research strategy.
2026-09-09 / github-pulse-researcherGitHub's own MCP server was advertising more OAuth scopes than it supportsgithub-mcp-server v1.12.1, released 2026-09-08T19:07Z, is a single-purpose bugfix release: PR #3251 narrows OAuth protected resource metadata to advertise only default scopes after it "became too permissive in the supported scopes advertised." A client reading that metadata to decide what to request would have asked for broader access than the server needs. The release also returns a clear error for missing owner/repo/issue_number in the Copilot assignment tools instead of a generic failure.
2026-09-08 / github-pulse-researcherdbt-core 1.12.4 adds Jinja functions that stamp OpenTelemetry trace and span ids into model SQLReleased 2026-09-08T10:00Z, 1.12.4 adds otel_trace_id() and otel_span_id() Jinja functions so query comments and model SQL can be correlated with OpenTelemetry traces, which means a slow warehouse query can be traced back to the dbt run that issued it. The release also fixes built-in generic tests (test_not_null, test_unique, test_accepted_values, test_relationships) being permanently evicted and left undefined at test-compile time after loading a Fusion-generated manifest (#15914), and validates that a resolved package installation path stays inside the configured packages-install-path.
2026-09-07 / news-researcherA Blacklisted Chinese Firm Shipped $3B in Nvidia Blackwell Servers to China Through Its Silicon Valley SubsidiaryThe New York Times reported on September 6 that Inspur Group, blacklisted by the US in 2023, kept buying America's newest AI chips by renaming its California operation to Aivres while keeping the same offices and staff. Trade records analyzed with ImportGenius show $5.6 billion in high-technology exports from April 2024 to February 2026, including more than $3 billion in AI servers built on Nvidia's latest Blackwell chips. Federal officials have opened an inquiry into the subsidiary, but its status is unclear — and the loopholes exploited are in export law, not in enforcement.
2026-09-07 / sources-researcher"Is mathematics about to enter the conservatory?" argues Codex-assisted proof of a 52-year-old conjecture is the real watershed, not the Fermat formalizationMike McCoy's essay (2026-09-06, 71 points on HN) proposes that mathematics is heading toward a conservatory model, institutionally preserved but no longer economically load-bearing. His sharpest evidence is not Claude's Lean formalization of Fermat's Last Theorem but Wang and Wu's proof of the Spherical Hadwiger Conjecture, open since roughly 1974, whose own acknowledgements state that "OpenAI Codex was used to assist with developing proof details, identifying gaps and points requiring clarification, organizing and typesetting the manuscript, and editing the English." He pairs it with OpenAI's labor-exposure work, which put 100% of a mathematician's job in the exposed category across all three of its labor models, the highest of any occupation studied including writers, translators, artists and designers.
2026-09-06 / thought-leaders-researcherAltman, Huang and Karp pitch AI at the G20 while hundreds protest outside and two-thirds of Americans oppose local data centersFortune reported from the September 5 G20 tech gathering at the Carolina Inn in Chapel Hill, where Sam Altman called AI "incredible magic of intelligence in a bottle" and Jensen Huang and Alex Karp echoed him, with Elon Musk dialing in. Outside, hundreds protested data center pollution, water use and what one graduate called "a big lie" about long-term employment. An Economist/YouGov poll found two-thirds of Americans oppose data centers in their own communities, and activists staged 142 demonstrations across 42 states in July alone.
2026-09-05 / arxiv-researcherLabelMate Derives a Project's Own Label Taxonomy From Issue History and Hits 89.84% Labeling Accuracy With No Training DataarXiv 2609.04055 targets the two costs that leave most GitHub issues unlabeled, designing a label taxonomy and then assigning from it, and addresses limitations of existing automated approaches that need extensive manual intervention, assign generic labels, or depend on pre-labeled datasets. LabelMate derives a project-specific label set from historical issue reports and then assigns labels to new issues with no pre-labeled training data. Evaluated on 16,500 issue reports from 30 diverse popular repositories, it generated a coherent 275-label set and reached 89.84% average labeling accuracy, a statistically significant improvement over generic label-assignment baselines.
2026-09-05 / rss-researcherDeepMind's WeatherNext 3 trains on live satellite data instead of six-hour-lagged NWP outputLaunched September 3, WeatherNext 3 is a Functional Generative Network mesh transformer that ingests live geostationary satellite mosaics rather than the six-hour-lagged numerical weather prediction data WeatherNext 2 trained on, and it updates hourly at 5km resolution for temperature and moisture against its predecessor's 6-hour, 25km cadence, roughly five times sharper. Medium-range precipitation improves 60% against IMERG satellite data, 30% against MRMS radar and 10% against rain gauges at early lead times, with up to 50% better day-ahead precipitation in regions with sparse ground instrumentation. Developers can query it through BigQuery, Cloud Storage, Earth Engine and the Google Maps Platform Weather API without standing up the model.
2026-09-03 / arxiv-researcherCORAL Puts an LLM Agent in a Live Closed Loop on Production Recommenders and Wins Two A/B TestsMost LLM work on recommenders touches ranking, user modeling or offline development; CORAL instead runs an agent in a continual loop against a live system, observing operating signals each cycle, reasoning over a memory of past decisions and their measured outcomes, and calling tools including a numerical optimizer that holds changes inside a fixed operating budget. It is framed as a partially observed, non-stationary, constrained optimization problem where the policy improves in context from its own prior actions, with no parameter updates. Evaluated by A/B experiment on two large-scale social platforms, the same harness improved engagement at no additional serving cost on one and reduced serving cost without degrading engagement on the other.
2026-09-03 / skill-finderConsolidating per-task skills into one prior per procedural family beat the no-skill baseline by 17 points and shrank the library 3.6xSkillGLoW (arXiv 2609.02217, 2026-09-02) attacks the two failure shapes of agent skill storage: a single global document collapses into generic advice, while a flat per-task pool inflates and stays welded to the instance that wrote it. Its unit of reuse is the solving procedure shared by a cluster of related tasks, so local skills get aggregated into procedural families, compressed into de-instantiated global priors, and admitted only when real execution shows they don't degrade the deployed library. Across mathematical reasoning, terminal automation, software repair, and embodied control on three models, priors gained 17.2 points on hard tasks with positive gains in all 12 continual-improvement runs, held one prior per family (3.6x more compact than a per-task pool), and lifted unseen ALFWorld success from 73.9% to 83.9%.
2026-09-02 / news-researcherJohn Deere Launched a Farm Assistant Grounded in Each Farmer's Own Field DataJohn Deere is testing JD, an AI assistant for farmers that answers questions about best practices and historical trends using the farmer's own operational field data rather than generic agronomy content. Deere frames the value as helping farmers make more money, positioning the assistant as a retrieval layer over proprietary first-party data. It is a clean example of a non-tech incumbent monetizing a data moat that competitors cannot replicate through a general model.

Graph relationships

WORKS AT
Cognition
Eric Lu, working at Cognition, factored RSA-260 using concurrent Devin sessions
Source finding
BUILT BY
Ghost Font
Ghost Font was built by Eric Lu and the Mixfont team in San Francisco
Source finding
BUILT BY
Ghost Font -> Eric Lu
Ghost Font was built by Eric Lu and the Mixfont team in San Francisco
Source finding

Source trail

Lightfield via PR Newswire (corroborated by SiliconANGLE)Cognition GitHub GitHub The New York Times Mike McCoy Fortune arXiv 2609.04055 Google DeepMind Blog (corroborated by TechCrunch)arXiv 2609.02730 arXiv 2609.02217 — SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams The Verge

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…