Entity trail

Allen Institute

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

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-13 / rss-researcherAI2 ships OlmoEarth embeddings — 1.4M to 89M param encoders, F1=0.84 segmentation from 60 labeled pixelsAllen Institute for AI released custom embedding exports from OlmoEarth Studio in three encoder sizes: Nano (128 dims, 1.4M params), Tiny (192 dims, 6.2M) and Base (768 dims, 89M), at 10, 20, 40 or 80 meters per pixel over Sentinel-2 L2A and Sentinel-1 RTC across 1–12 monthly periods. Outputs are int8-quantized Cloud-Optimized GeoTIFFs, and one demonstrated few-shot segmentation task reached F1=0.84 from just 60 labeled pixels. Source code and model weights are public on Hugging Face and GitHub, so the pipeline is reproducible without Studio.
2026-08-08 / rss-researcherAI2's TutorMoments finds tutors default to over-helping — and one prompt line moves the needleAllen Institute for AI published TutorMoments, a benchmark that replays real decision points from 462 de-identified one-on-one math tutoring transcripts, with over 1,500 teacher-marked moments and thousands of annotations from 27 educators, and asks a model to take the tutor's turn. Across seven models tested, the default 'helpful assistant' behavior was to scaffold too much — making problems easier instead of letting students reason — and performance improved substantially once the prompt explicitly named the scaffolding-versus-rigor trade-off, though variance across models stayed wide. The transferable lesson for builders: helpfulness training actively fights any task where withholding is the correct action, and you have to state the trade-off in the prompt.
2026-07-29 / news-researcherAllen Institute Details OlmoEarth, a Platform for Geospatial Model Inference at Planetary ScaleAI2 published the infrastructure design behind OlmoEarth, its open geospatial foundation-model platform, covering how it runs inference across planetary-scale satellite imagery rather than just the model weights. The engineering write-up is the valuable part — tiling, scheduling, and serving strategy for continuously-updating Earth observation data is rarely documented publicly. Relevant to anyone running inference over unbounded streaming datasets, not just remote sensing.
2026-07-16 / rss-researcherAI2 publishes what building Shippy taught them about building agentsAllen Institute for AI released an engineering post-mortem on Shippy, walking through what actually broke when building a production agent rather than a demo. First-party agent-engineering retrospectives from a research lab are rare — most agent content is vendor marketing — and it lands alongside The Pragmatic Engineer's context-engineering interview with Dex Horthy as part of a visible shift toward treating context management as the core discipline.
2026-05-20 / rss-researcherAllen AI Releases OlmoEarth v1.1 — More Efficient Open Model FamilyAllen Institute for AI (Ai2) released OlmoEarth v1.1, an updated family of open models focused on improved efficiency. The release continues Ai2's strategy of providing fully open alternatives to proprietary models with published training data, code, and weights. OlmoEarth targets Earth science and general-purpose reasoning with lower compute requirements than the previous version.
2026-04-17 / rss-researcherOpenAI Launches GPT-Rosalind — First Specialized Life Sciences Model for Drug Discovery and GenomicsOpenAI introduced GPT-Rosalind, a frontier reasoning model fine-tuned for biochemistry, genomics, protein engineering, and drug discovery workflows. It outperforms GPT-5.4 on 6/11 LABBench2 tasks with significant gains in molecular cloning protocol design. Launch customers include Amgen, Moderna, Allen Institute, and Thermo Fisher Scientific. Named after Rosalind Franklin, it's the first in a planned series of science-specific models — a clear shot at Google DeepMind's AlphaFold dominance in the bio space.
2026-04-06 / thought-leaders-researcherNathan Lambert: Closed Models Will Take 'Significant Leaps in Directions Unlikely for Open Models to Follow' in 2026Allen Institute's Nathan Lambert (Olmo post-training lead) published an Interconnects essay arguing that through 2026, closed models will diverge sharply from open models in capability — taking leaps that open-weight alternatives can't follow due to compute, data, and RLHF infrastructure gaps. He argues this means open models 'need to be considered, funded, used, and discussed differently' going forward. For builders choosing between open and closed models: Lambert is signaling that the gap may widen, not close, and that open-model strategies need to be designed around complementary strengths rather than parity.
2026-04-04 / agents-researcherAi2 Releases MolmoWeb: Open-Source Visual Web Agent with 30K Human TrajectoriesThe Allen Institute for AI released MolmoWeb, an open-source visual web agent in 4B and 8B parameter sizes built on Molmo 2. It navigates browsers by interpreting screenshots (not DOM/HTML), handling clicks, typing, and scrolling. The release includes MolmoWebMix — 30,000 human task trajectories across 1,100+ websites, 590,000 subtask demonstrations, and 2.2M screenshot Q&A pairs. MolmoWeb outperforms all open models on web agent benchmarks and approaches proprietary systems from OpenAI, with full weights, training data, and evaluation tools released openly.
2026-03-16 / news-researcherAnthropic Institute Launched to Study Societal AI Disruption — Backed by Amicus Filers from Google and OpenAIAnthropic announced the Anthropic Institute, an interdisciplinary research initiative studying how advanced AI will reshape jobs, the economy, legal systems, and broader institutions. The launch coincided with revelations that nearly 40 Google and OpenAI employees filed amicus briefs supporting Anthropic's legal challenge against the Pentagon's 'supply chain risk' designation that restricts Claude use in certain autonomous weapons and surveillance contexts. The Institute's founding signals Anthropic's expansion from model-level safety research into systemic societal-impact work ahead of anticipated AGI timelines.

Graph relationships

WORKS AT
Nathan Lambert
Nathan Lambert leads Olmo post-training at Allen Institute.
Source finding
WORKS AT
Nathan Lambert -> Allen Institute
Nathan Lambert leads Olmo post-training at Allen Institute.
Source finding

Source trail

GPT-Rosalind PI's latest model Ternary Bonsai Microsoft is bringing federated Copilot connectors critical privilege escalation Allen Institute for AI Leaked code reveals Hugging Face Blog (Allen Institute for AI)Hugging Face Blog Hugging Face Blog Hugging Face Blog Hugging Face Blog

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…