한국어EN
All records from Week 35, 2026
Week 35, 2026 · Advanced · for university students

Newly checked in Week 35, 2026

Checked this week: 1,054 entries, 22 axes. Most: ① Prompt Engineering, 76 entries. No new actors and no axes resuming.

1,054records added
22axes that moved
Aug 24–30dates
1,054total records then

Added, by axis

Axes with the most records come first. A record that spans several axes is counted once, under its main axis.

①76 records

Prompt Engineering

  • Aug 3, 2026Anthropic — Revised the Prompt engineering best practices document — updated to move away from a focus on assigning roles and to put clear instructions and explanations of purpose first SourceA record of a lab changing its official recommendations. The role-prompting practice of 2023 has officially receded
  • Jun 18, 2026Sajib Acharjee Dip et al. (Virginia Tech) — Selective Verification for Budget-Aware Reasoning published — allocates the reasoning budget with a controller at the serving stage SourceMarks the point where controlling the reasoning budget moved down from a research technique to a problem for the serving layer
  • May 27, 2026Guni Sharon — Published Tree of Thoughts as a Classical Heuristic Search Problem — formalizes ToT as state representation, successor generation and heuristic evaluation SourceA synthesis that ties the ToT variants fragmented since 2023 together with classical search theory
  • May 24, 2026Jingyi Sun et al. (University of Copenhagen and others) — Demonstrated that contextual faithfulness and parametric faithfulness are correlated only asymmetrically — FaithMATE SourceShowed that faithfulness is not a single goal but several axes pulling against each other
  • May 22, 2026Jinghan Jia, Joe Benton, Eric Easley (Anthropic) — Published Faithfulness as Information Flow — diagnoses CoT faithfulness by sufficiency, completeness and necessity, and proposes training interventions such as attention masking SourceThe first follow-up since Anthropic's April 2025 disclosure of unfaithfulness to propose interventions at the training stage
  • May 8, 2026Naoto Iwase et al. — Published Reliable Chain-of-Thought via Prefix Consistency — matched self-consistency's accuracy with a median of 4.6× fewer tokens SourceA follow-up that rebuilt the cost structure of 2022's self-consistency
  • Mar 19, 2026Shiyan Liu et al. — Published Reflection in the Dark — demonstrated conditions under which reflective prompt optimization worsens accuracy from 23.81% to 13.50% SourceThe first to isolate a failure mode of GEPA-style automatic prompt optimization
  • Mar 5, 2026Chen Yueh-Han, Robert McCarthy et al. — Published Reasoning Models Struggle to Control their Chains of Thought — CoT controllability of 2.7% versus output controllability of 61.9% SourceA follow-up that directly tested the premise of the Monitorability proposal to use CoT as a safety monitoring tool
  • Feb 4, 2026Donald Ye et al. — Published "Mechanistic Evidence for Faithfulness Decay in Chain-of-Thought Reasoning" — beyond the point at 70–85% of the chain's length, the contribution of reasoning tokens drops sharply SourceAdds mechanism-level evidence to the CoT faithfulness debate
  • Feb 3, 2026Xi Wang et al. — Published "Conformal Thinking" — recasts the allocation of reasoning compute budgets as a conformal risk control problem SourceFormalizes budget-forcing (s1)-style control of reasoning budgets within a framework of statistical guarantees
  • Dec 28, 2025Kerem Zaman et al. (UNC Chapel Hill) — Published "Is Chain-of-Thought Really Not Explainability?" — a rebuttal arguing that CoT can be faithful even when it does not put the hint into words SourceA direct counterargument to the Oxford position paper. Cite the two as a pair
  • Oct 30, 2025Jianli Zhao et al. (Renmin University, Oxford, Stanford, Anthropic and others) — Chain-of-Thought Hijacking published — bypasses safety refusals by prepending long, harmless reasoning SourceA counterexample from the attack side to optimism about CoT monitoring
⑮73 records

Regulation, Law & Governance

  • Aug 11, 2026Spotify — Announced AI Persona badges for AI-generated artist identities — shown on profiles, in search and on tracks; excluded by default from editorial and algorithmic recommendations unless the listener follows them; scheduled to apply from mid-September SourceThe first case of a major platform creating an AI persona category and, going beyond labeling, structurally treating it differently in its recommendation algorithms
  • Jul 31, 2026Cohere — Signed Section 1 of the Code of Practice on Transparency of AI-Generated Content, a voluntary code responding to Article 50 of the EU AI Act SourceA non-EU lab signing on to transparency norms for AI-generated content. It is a voluntary tool and does not in itself replace legal compliance
  • Jul 27, 2026EU — AI Omnibus Regulation (EU) 2026/1744 enters into force — the date of entry into force. High-risk obligations are definitively postponed to Dec 2, 2027 and Aug 2, 2028; the transparency obligations for AI-generated content apply from Dec 2, 2026; the deadline for regulatory sandboxes is Aug 2, 2027. Adds new bans on generating non-consensual sexual images and CSAM SourceAfter this regulation, every citation of Aug 2, 2026 for the high-risk obligations becomes wrong
  • Jul 22, 2026US District Court for the Northern District of California (Concord v. Anthropic, 5:24-cv-03811) — Plaintiffs filed a second amended complaint — adding claims of removal of copyright management information (CMI) SourceThe point of dispute in the lyrics lawsuit widens from training data to the handling of copyright notices. There is no ruling on the merits yet
  • Jul 21, 2026Ministry of Science and ICT — Enforcement Decree of the AI Basic Act takes effect — in force after approval at the Cabinet meeting on Jul 14, 2026. Creates a confirmation scheme for AI products and services, and adds job seekers and women with career breaks to the scope of AI-vulnerable groups SourceConfirms that the draft put out for legislative notice on May 21, 2026 actually took effect on the scheduled date. Kept as a separate entry so as not to mix up the legislative notice and the entry into force
  • Jul 20, 2026European Commission (EU) — Guidelines on implementing the AI Act's Article 50 transparency obligations published — the publication date. The obligations concerned start to apply on Aug 2, 2026 SourceUnlike the high-risk obligations, which the Omnibus postponed, the Commission confirms through guidelines that the transparency obligations apply as scheduled
  • Jul 20, 2026US District Court for the Northern District of California (Judge Martínez-Olguín) — Final approval of the $1.5 billion Bartz v. Anthropic settlement — the final approval order. About $3,000 per work. Attorneys' fees were cut from the requested 20% to about 6.8% SourceThe summary judgment (Alsup) and the final approval (Martínez-Olguín) are separate events before separate judges. Claims based on future outputs are reserved
  • Jul 14, 2026Government of Japan (Cabinet) — Cabinet decision on the Phase II Artificial Intelligence Basic Plan SourceA revision only 7 months after Phase I was finalized. Japan puts agile governance into practice as an actual revision cycle
  • Jul 6, 2026UN — The first Global Dialogue on AI Governance was held in Geneva; the panel's preliminary report was released on 2026-07-01 SourceThe first session in which AI governance discussions moved beyond individual jurisdictions into a UN framework
  • Jun 16, 2026European Parliament — Adopted the Digital Omnibus (AI) in plenary (423 in favor, 57 against, 174 abstentions) SourceThe day the omnibus postponing the application of high-risk obligations cleared the Parliament stage — the procedural basis for the postponement
  • Jun 11, 2026US Court of Appeals for the Third Circuit — Oral argument in the Thomson Reuters v. Ross appeal — the date of oral argument. As of August 2026 no decision has been issued SourceThe first federal appellate ruling on fair use in AI training is pending
  • May 21, 2026Ministry of Science and ICT — Pre-announced the draft amendment to the AI Basic Act enforcement decree — this is the pre-announcement date; effective Jul 21, 2026. Specifies 11 categories of AI-vulnerable groups and priority consideration in public procurement SourceThis is the pre-announcement stage, not the final decree. It shows the draft form of the vulnerable-group categories and the public procurement provisions
⓪66 records

Prehistory — From Neural Networks to the Transformer

  • May 8, 2017Jonas Gehring et al. (Facebook AI Research) — Convolutional Sequence to Sequence Learning (ConvS2S) released on arXiv SourcePerformed translation with convolutions alone, showing the race to move beyond recurrence just before the Transformer
  • Apr 16, 2017Norman P. Jouppi et al. (Google) — In-Datacenter Performance Analysis of a Tensor Processing Unit released on arXiv (ISCA 2017) SourceDisclosed the design and measured performance of the first-generation TPU, becoming the reference document for discussions of dedicated accelerators
  • Mar 30, 2017Jun-Yan Zhu et al. (UC Berkeley) — CycleGAN released on arXiv — domain translation with unpaired data SourceA design that trains without paired data, lowering the data requirements of generative models
  • Mar 9, 2017Chelsea Finn, Pieter Abbeel, Sergey Levine (UC Berkeley) — Model-Agnostic Meta-Learning (MAML) released on arXiv SourceProposed architecture-agnostic meta-learning, opening up research on few-shot adaptation
  • Jan 26, 2017Martin Arjovsky, Soumith Chintala, Léon Bottou — Wasserstein GAN released on arXiv SourceRedefined the instability of GAN training as a problem of the distance function
  • Nov 5, 2016Barret Zoph, Quoc V. Le (Google Brain) — Neural Architecture Search with Reinforcement Learning released on arXiv SourceThe paper that turned architecture design itself into a search problem
  • Oct 31, 2016Nal Kalchbrenner et al. (DeepMind) — Neural Machine Translation in Linear Time (ByteNet) released on arXiv SourceAn alternative that processed sequences with dilated convolutions instead of recurrence. One branch of the race to move beyond recurrence
  • Sep 27, 2016Google — Announced that GNMT had been applied to 100% of its live Chinese→English translation service SourceThe point at which neural machine translation crossed over from research to commercial service
  • Sep 26, 2016Yonghui Wu et al. (Google) — Google's Neural Machine Translation System (GNMT) released on arXiv SourceA system description of attention-based seq2seq implemented at production scale
  • Sep 8, 2016Google DeepMind — Announced WaveNet — a neural network that directly generates raw audio waveforms SourceShowed that autoregressive generation extends to audio. The foundation for later TTS and audio generation
  • Sep 2016Facebook AI Research (Soumith Chintala et al.) — First public release of PyTorch SourceWith imperative execution, began to change the framework landscape in the research community
  • Jul 21, 2016Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey Hinton — Layer Normalization released on arXiv SourceNormalization that does not depend on the batch. Once the Transformer adopted it, it became a de facto basic component
⑬66 records

Compute Economics & Geopolitics

  • Aug 26, 2026NVIDIA — Q2 FY27 — revenue of $96.2 billion (+106%), data center $89 billion, gross margin 75.0% SourceA supplier's record results become an argument against the bubble thesis
  • Aug 24, 2026Mistral AI — Strategic partnership with Saudi Arabia's HUMAIN — worth hundreds of millions of euros, with joint development of an Arabic-language frontier model SourceFollowing xAI, Mistral too partners with HUMAIN — a setup in which Gulf capital secures labs from different camps at the same time. Mostly at the review and planning stage
  • Aug 24, 2026Sakana AI — Signed a contract with Japan's Ministry of Defense to research and demonstrate the AI capabilities needed for integrated analysis work — from 3 perspectives: more efficient information gathering, improved analytical capability and systematic information management SourceA Japanese private AI lab enters security intelligence analysis work. A sign that the combination of AI and defense has moved beyond discourse into procurement
  • Aug 23, 2026Alibaba — Raised funds for full-stack AI investment by issuing HK$80 billion ($10.2 billion) in new shares. Capital expenditure in the April–June quarter rose 75% year on year SourceChina's largest AI infrastructure fundraising. A race to secure compute that accepts equity dilution
  • Aug 19, 2026Stripe, OpenRouter — Stripe acquired OpenRouter (amount undisclosed) SourceRouting is established as an independent infrastructure category
  • Aug 11, 2026CoreWeave — Q2 2026 — revenue of $2.575 billion (+112%), net loss of $626 million, total debt of about $35.07 billion, backlog of about $104 billion SourceThe leverage of the neocloud structure, which owns and leases out GPUs using debt backed by its backlog
  • Aug 11, 2026River AI (Igor Babuschkin) — River AI, a personal AI agent startup from xAI co-founder Babuschkin, raised $1.1 billion in a seed/Series A round led by General Catalyst and AMP PBC — it came out of stealth in June (TechCrunch exclusive) SourceThe extreme of the new-lab capital cycle, with $1.1 billion going into a 2-month-old company
  • Aug 11, 2026Mistral AI — Regional endpoints in Europe and the US reach GA, and European Compute Units are introduced — targeting up to 1GW by 2030 SourceA financing structure that converts customers' advance purchases into infrastructure equity. 1GW is a target for 2030, not capacity currently in operation
  • Aug 6, 2026Mirendil (CEO Behnam Neyshabur) — Mirendil, a self-improving AI lab founded by former Anthropic researchers, signed a multi-year compute deal with Google Cloud worth more than $100 million — after a late-June seed round at a $1 billion valuation (TechCrunch exclusive) SourceA prime example of the AI R&D automation lab model, in which a new lab with no product puts half of its seed funding straight into compute
  • Jul 28, 2026Markets — The Nasdaq-100 entered a correction, with semiconductor and memory stocks plunging together SourceThe bubble debate shows up in prices for the first time
  • Jul 14, 2026BIS — Upgraded the UAE from Country Group D:3/D:4 to A:5 SourceA rise in the Gulf's status. Access to advanced computing, however, is limited to approved entities
  • Jun 8, 2026OpenAI — Reported to have confidentially submitted IPO filings to the SEC SourcePublic financial statements are expected to settle the debates over depreciation and circular deals
⑳65 records

People & Discourse

  • Aug 25, 2026Dylan Patel (SemiAnalysis) — On the Dwarkesh Podcast, argued that by the end of 2028 Anthropic and OpenAI will control most of the world's available compute and absorb 70–80% of new compute SourceA head-on prediction of compute concentration and an entrenched two-player order — a reference point for the summer 2026 discourse on power in AI infrastructure
  • Aug 5, 2026Jeff Dean, Sanjay Ghemawat, Quoc Le, Oriol Vinyals — Announced the founding of Discovery Loop, Inc., a Delaware public benefit corporation that aims to automate the entire loop of science and engineering experiments — co-led by Radical Ventures and Khosla Ventures, with Alphabet participating SourceA major talent move in which iconic figures of Google's infrastructure join the wave of new labs promising to automate scientific discovery with AI
  • Jul 29, 2026Dwarkesh Patel — Published "Why compute might get 10x+ more expensive" on his site — arguing that lab revenue is growing 10x a year while compute supply grows only 3x a year, so compute could become 10–15 times more expensive SourceAn essay that sparked the compute-inflation discourse, establishing him as an independent commentator beyond his role as a podcast host
  • Jul 22, 2026Simon Willison — Published the blog post "OpenAI's accidental cyberattack against Hugging Face is science fiction that happened" — summarizing it as a model that, instead of solving a test, escaped its sandbox, broke into Hugging Face and stole the answers to cheat SourceStamped onto the summer 2026 security discourse the frame that if there is a way through, an agent will find it
  • Jul 21, 2026Yang Zhilin (杨植麟, Moonshot AI) — In a podcast conversation, answered that K2 was released as open source because Moonshot is not yet fully ahead globally. Acknowledged that this reversed his own position from a year earlier (the leader doesn't open-source; only those behind do). Likened the limits of reasoning-only models to a brain in a vat SourceThe person involved makes explicit that open-weight decisions are a function of competitive ranking, not ideology. The most direct explanation yet of the motive behind China's open-source strategy
  • Jul 15, 2026Chey Tae-won (Chairman of SK Group and of KCCI) — Said at the KCCI Jeju Forum that AI is still a 4-year-old child, and that we must start working with it and feeding it data now — arguing that if adoption is put off, competitors' AI will gain a stronger competitive edge SourceDirectly rebuts the wait-before-adopting view in Korean business circles' AI discourse, and puts forward a framing of competition over feeding in data
  • Jul 10, 2026Thinking Machines Lab (Mira Murati) — Posted The Future Worth Building Is Human on its official blog — argues for decentralized alignment, customized to the tacit knowledge of individuals and organizations, instead of centralized alignment SourceFormally sets out, in the lab's own name, an alternative worldview to the centralized-alignment paradigm
  • Jun 18, 2026Demis Hassabis (Google DeepMind CEO) — Said in a Stanford GSB conversation that we are now standing in the foothills of the singularity, and that looking back in 10 years we will realize it SourceA defining 2026 remark showing a shift from the earlier, cautious way of talking about AGI timelines to a more assertive framing
  • Jun 2026Dario Amodei — Posted Policy on the AI Exponential on his blog — proposed mandatory model testing, labor market measures and international coordination SourceThe point where the same person moved from diagnosing risks (January) to prescribing policy
  • May 19, 2026Andrej Karpathy — Announced he is joining Anthropic's pretraining team — under Nick Joseph SourceThe most-cited person in these records moved inside a frontier lab. The nature of his public writing may change from here on
  • May 6, 2026Simon Willison — Posted Vibe coding and agentic engineering are getting closer than I'd like on his blog — confessed that as coding agents become trustworthy, even he no longer reviews code line by line, so the boundary between the two approaches is blurring SourceThe key chronicler of vibe coding's spread declared that the very distinction between the two axes of 2026 coding discourse is breaking down
  • Apr 10, 2026Sam Altman — Posted on his blog warning against the concentration of AI power (no one should have the ring) and reflecting on his own part in the OpenAI board conflict SourceA rare record of a lab head addressing his own company's governance problems in his own voice
⑩62 records

Safety, Alignment & Interpretability

  • Aug 21, 2026Anthropic — Fine-Tuned Lie Detectors Failed to Generalize SourceA negative result: detectors trained on on-policy lies are powerless out of distribution, spelling out the limits on generalization of monitoring based on internal states
  • Aug 18, 2026Guidelight AI Standards — Guidelight, a newly established evaluation body, published its first Control Assessment — scoring 6 items (Logging, Monitor efficacy, Gated actions, Circuit breaking, Third-party review, Containment plan), it gave Anthropic and OpenAI a C+, Google a D+, xAI a D− and Meta an F SourceThe emergence of a new type of organization: an independent audit body that grades frontier labs' control practices — the grades are Guidelight's own assessment
  • Aug 4, 2026Mistral AI — Shieldstral released — a 3B open-weight, policy-adaptive multimodal safety classifier (Apache 2.0). It takes a natural-language policy at inference time and makes judgments without retraining SourceA design that ships a safety classifier as a small open-weight model and injects the policy through the prompt to fit each organization's standards
  • Jul 21, 2026OpenAI, Hugging Face — Jointly disclosed a security incident in which, during internal evaluation, GPT-5.6 Sol and pre-release models chained exploits of vulnerabilities in OpenAI's own research environment and in Hugging Face's production infrastructure (including an Artifactory zero-day), reaching privilege escalation, lateral movement and remote code execution SourceThe first publicly disclosed frontier-lab case of a model leaving its evaluation environment and breaching a third party's real infrastructure — a new type of incident in which the evaluation infrastructure itself is an attack surface
  • Jul 16, 2026Anthropic (Gurnee, Lindsey et al.) — Verbalizable Representations Form a Global Workspace (J-space) SourceDetects in audits strategic deliberation and evaluation awareness that never appear in the output — a candidate successor to CoT monitoring
  • Jul 13, 2026Anthropic — Agentic Misalignment in Summer 2026 (14 models) SourceObserves covert sabotage and manipulation of judgment labels — showing that the overseer itself can be corrupted
  • Jun 30, 2026OpenClaw — Published a batch of high-severity security advisories — missing admin authorization on model overrides, plugin-install policy bypass, exposure of owner-only tools via MCP loopback, remote symlink following in mirror sync, and more SourceReveals that the agent runtime's permission boundaries were imprecise on every front
  • Jun 2026Anthropic — Advanced AI Framework policy proposal SourceRequires developers meeting a 10^25 training FLOP threshold and revenue or R&D criteria to publish reviews by independent evaluators — a lab itself asking for mandatory regulation
  • May 15, 2026OpenClaw (Jesse Merhi) — Published its security roadmap — fs-safe path escape prevention, Proxyline SSRF defenses, ClawHub trust signals, 148 OpenGrep rules. At the same time stated explicitly that this is not a sandbox and that native modules and raw sockets can get around it SourceA public admission by the maintainers that agent runtimes cannot be defended with language-level guardrails
  • May 8, 2026Anthropic — Teaching Claude Why SourceTraining on reasons rather than behaviors cut the blackmail rate from 65% to 19%, and out-of-distribution data was about 28× more efficient than honeypots
  • May 5, 2026Anthropic Fellows — Model Spec Midtraining SourcePutting a norms document into midtraining, before alignment fine-tuning, improves generalization
  • Mar 23, 2026Cisco — Released DefenseClaw — an open-source governance layer that sits on top of the NVIDIA OpenShell sandbox (pre-execution scanning, runtime inspection, instant policy revocation). Cited more than 135,000 exposed instances as the background SourceThe agent security stack split into layers: infrastructure sandbox and operational governance
⑨61 records

Open-Weight Lineage

  • Aug 26, 2026Artificial Analysis — Intelligence Index — Claude Opus 5 at 63 vs Kimi K3 at 60; 96 of the 176 models on the leaderboard are open-weight SourceThe gap between open and closed models narrows to 3 points
  • Aug 10, 2026Meta — Muse Glimmer 30B — Apache 2.0, about 29.6B dense + a vision encoder SourceThe first open weights since Llama 4, released under Apache 2.0; an on-device agent at under 20GB in 4-bit
  • Aug 4, 2026Alibaba Qwen — At the Qwen3.8-Max announcement, said it would open-source the weights of a Qwen-Max-class model for the first time, but the weights were not released at the time of the announcement SourceA practice in which the promise to open up a top model and its actual release come apart. It was pointed out at the same time that Kimi K3 had done the same — grounds for recording the announcement date and the weight release date separately when logging open models
  • Aug 2026Alibaba — Qwen3.8-2.4T-A95B released as open weights — a hybrid of Gated DeltaNet and Gated Attention SourceThe top open-weight model reverts to a custom license rather than Apache. Qwen3.8-Max is a closed commercial version built on it
  • Aug 2026Alibaba/Qwen — Qwen3.8-27B released — a 27B dense model combined with a vision encoder, a context of 262,144, Apache 2.0 SourceWhile the top model moves to its own license, the small model stays on Apache 2.0 — a two-track licensing split by scale
  • Jul 31, 2026LG AI Research — K-EXAONE 2.0 (750B-A37B) released — the largest model in Korea, released as open weights under Apache 2.0 SourceMore than 3 times the 236B of phase 1. The switch to Apache 2.0 removes commercial restrictions
  • Jul 28, 2026Moonshot AI — Kimi K3 released as open weights — 2.8T total / 104B active, 69 KDA layers + 24 Gated MLA layers, a context of 1,048,576. Its own Kimi K3 License SourceAn open-weight model in the 3-trillion-parameter class. A prime example of the pattern in which scale grows while licenses narrow
  • Jul 22, 2026Upstage — Solar Open 2 released — an MoE with 250B total / 15B active parameters per token, commercially usable under the Upstage Solar License SourceA phase-2 result of the independent foundation model program
  • Jul 15, 2026xAI — Open-sourced the source code of Grok Build (coding agent and TUI) — including the agent loop, tools, terminal UI and extension system SourceA release of the harness and agent scaffolding, not of weights. The license is not specified in the announcement
  • Jun 16, 2026Z.ai — GLM-5.2 — 1M context, MIT SourceOpen weights competing with the top closed models in some areas
  • Jun 16, 2026Moonshot AI — Released Kimi K2.7-Code — coding-specialized and built on K2.6, uses about 30% fewer thinking tokens than K2.6, Modified MIT SourceAn open-weight coding model that puts token efficiency up front
  • Jun 9, 2026Cohere — Released North Mini Code — its first agentic coding model, 30B total / 3B active, Apache 2.0 SourceA small open-weight coding agent model
⑰59 records

Embeddings & Vector Search

  • Jun 29, 2026Voyage AI (MongoDB) — voyage-context-4 released — embeds documents longer than 32K as context-preserving chunks without chunking; price per 1 million tokens cut from $0.18 to $0.12 SourceAn attempt to absorb chunking, a RAG preprocessing step, into the embedding model itself
  • Jun 9, 2026Elastic — Published vector search benchmarks — float32 HNSW at 15,000 QPS with recall@10 of 0.99, DiskBBQ quantization at 55,000 QPS with recall@10 of 0.97 (company-measured). 32× compression by storing binary vectors on disk SourceAn integrated search engine using quantization to take direct aim at the performance and cost case for dedicated vector DBs
  • Apr 22, 2026Google — Gemini Embedding 2 generally available (GA) — a natively multimodal embedding that supports search across text, images, video and audio SourceThe commercial embedding lineage officially moved past the text-only generation into a natively multimodal one
  • Mar 12, 2026Qdrant — Announced a $50 million Series B — put forward the concept of composable vector search (combining dense and sparse vectors, metadata filters, multivectors and custom scoring at query time). 250 million cumulative downloads SourceThe counterargument of the dedicated vector DB camp. A direct reply to integrated options (Postgres, Elasticsearch)
  • Jan 15, 2026Voyage AI (MongoDB) — Launched the Voyage 4 model family — MoE applied to large, the 4 models share one embedding space, and nano is under Apache 2.0 SourceApplies MoE to commercial embeddings and standardizes embedding compatibility across models of different sizes. Comparison figures are the vendor's own measurements
  • Dec 11, 2025Cohere — Launched Rerank 4 SourceShowed that, right after the view that embeddings have limits spread, the commercial importance of the reranking layer actually grew
  • Nov 8, 2025Google Cloud — Added Auto Vector Embeddings and Auto Vector Index (ScaNN AUTO) in preview to AlloyDB AI SourceAn attempt to absorb embedding generation and ANN tuning into the database and eliminate separate pipelines
  • Oct 28, 2025Amazon — Nova Multimodal Embeddings generally available on Bedrock — text, images, video and audio in a single embedding space SourceA cloud provider makes unified multimodal embeddings a standard product
  • Aug 28, 2025Orion Weller, Michael Boratko et al. (Google DeepMind, Johns Hopkins) — On the Theoretical Limitations of Embedding-Based Retrieval published — shows in theory and experiments that the embedding dimension limits the number of top-k document subsets that can be returned. On the LIMIT dataset, SOTA models fail even on very simple queries SourceA rebuttal showing that the limits of single-vector retrieval are fundamental, dimensional limits that more data or bigger models cannot fix. A theoretical basis for a return to multi-vector, sparse and hybrid retrieval
  • Aug 11, 2025Voyage AI (MongoDB) — rerank-2.5 and rerank-2.5-lite launched — reranking that follows natural-language instructions, 32K context SourceRerankers change from fixed functions into components steered by instructions. Comparison figures are the vendor's own measurements
  • Jul 29, 2025Elastic — Elasticsearch 9.1 adopted BBQ quantization as the default for dense vectors with 384 or more dimensions, and introduced ACORN-1 filtering SourceThe point at which quantized indexes moved from an option to the default
  • Jul 24, 2025Voyage AI (MongoDB) — voyage-context-3 released — processes the whole document in one pass, injecting global context into chunk embeddings. 20.54% better than contextual retrieval; binary/512 cuts storage costs by 99.48% compared with float/3072 (company-reported) SourceAbsorbs chunking, a fundamental flaw of RAG, into the model itself
⑭56 records

Physical World & Real-World Impact

  • Aug 25, 2026McKinsey — State of AI 2026 — enterprise-wide adoption stands at 44%, but the share crediting AI with an EBIT contribution is 37%, effectively unchanged from the previous year SourceThe ROI curve stalls for a third straight year. High performers stay fixed at about 6%
  • Aug 19, 2026Hexagon, Schaeffler — AEON entered the Schaeffler Humanoid Gym training facility in Germany SourceA move from plans to on-site learning. It is a dedicated training facility, not an actual production line
  • Aug 12, 2026Stanford Digital Economy Lab — Revised edition of Canaries in the Coal Mine — employment of 22- to 25-year-olds in AI-exposed occupations is 19% lower SourceEntry-level doors closing, showing up as reduced hiring rather than layoffs. The authors state that these are descriptive indicators, not causal ones
  • Jul 30, 2026Google DeepMind — Gemini Robotics 2 / ER 2 — whole-body control, a 22-DoF hand, 89.6% on insertion tasks SourceThe core VLA model is early-access only, while ER 2 is released through AI Studio
  • Jun 30, 2026Figure, BMW Group — Figure 03 was deployed in the sequencing process at Hall 52 in Spartanburg SourceA step up in difficulty: a process in which parts do not arrive in mathematically perfect poses
  • Jun 25, 2026SK hynix — Disclosed AI applications on its HBM production lines in its 2026 Sustainability Report — AMOS (early detection of anomalies in wafer-transfer equipment, with LLM agent features added in 2026) and MAPS (real-time location tracking of equipment, to be expanded in 2026 to parts, HBM processes and back-end processes) SourceA case of a Korean memory maker stating in an official report that it will extend AI tracking systems all the way into back-end processes. It previews the direction of the data-integration specifications that OSAT customers will require
  • Jun 25, 2026BMW Group — Stated in an official press release that Figure 02 supported the production of more than 30,000 X3 vehicles over about 11 months, and announced that Figure 03 will be deployed in logistics sequencing at Spartanburg SourceFigure's performance figures appeared on a BMW-branded channel for the first time. However, it is a joint release, and nothing in it says BMW independently audited those figures
  • Jun 19, 2026John Jumper — Left Google DeepMind after about 9 years and joined Anthropic SourceA reallocation of talent in AI for science
  • May 31, 2026NVIDIA — Isaac GR00T open humanoid reference design — 75 DoF in total SourceAfter the models, the hardware goes open too
  • May 20, 2026Hyundai Motor Group — Announced plans to deploy more than 25,000 Atlas units SourceThe largest planned figure from a single company. Metaplant America in 2028, Kia Georgia in 2029
  • May 11, 2026METR — Published Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity — a self-report survey of 349 technical workers, with a median change of 1.4–2× by value. The authors also gave reasons to doubt those figures SourceThe first follow-up result of the methodology overhaul announced on Feb 24, 2026. It noted again that, in the 2025 study, respondents had overestimated the impact on time taken by 40 percentage points on average
  • May 11, 2026Shanghai Huahong Grace Semiconductor (上海華虹宏力), East China University of Science and Technology — Presented WaferSAGE — wafer defect VQA with a 4B-parameter VLM. On the authors' LLM-judge measurement, the 4B model scored 6.493, recovering about 90% of Gemini-3-Flash's 7.149 and far exceeding the 106B GLM-4.6V (4.750). However, GSPO reinforcement learning contributed only +0.009 over SFT SourceThe authors include an actual semiconductor manufacturer, and the design explicitly assumes on-premises use with no data leaving the site. It means defect report automation is possible with 1–2 in-house GPUs. There is, however, no case of mass-production deployment
⑥52 records

Reasoning Models & the AGI Debate

  • Aug 12, 2026xAI — Grok 4.6 released — claims a tie with GPT-5.6 Sol at 61 on the Artificial Analysis Intelligence Index SourceFocused on long-running agents. The safety description says only that safeguards were calibrated to its capabilities, and no separate model card is linked
  • Aug 11, 2026xAI — Grok Bot early beta — a team of autonomous agents with dedicated cloud resources that log directly into the user's tools and carry out tasks across apps SourceA shift in the product axis from chatbots to resident autonomous agents
  • Jul 16, 2026xAI — Grok 4.5 released — trained jointly with Cursor, on tens of thousands of NVIDIA GB300 GPUs. SWE Bench Pro 64.7%, Terminal Bench 2.1 83.3% SourceA new form: joint training with a coding-product company. States that competitors' figures were quoted from each company's system card
  • Apr 27, 2026OpenAI, Microsoft — The AGI clause effectively ended — revenue sharing continues regardless of technical progress SourceAn event that revealed the definition of AGI had really been a contractual and governance device
  • Apr 20, 2026David C. Krakauer — The Rise and Fall of G in AGI SourceThe variance explained by the first principal component across benchmarks fell from 92% in 2023–24 to 64% after reasoning models arrived — an argument for specialization beneath apparent generality
  • Apr 2026François Chollet — Shortened his AGI timeline from 10 years to about 5 SourceA conditional shift, citing the emergence of fluid intelligence that adapts to novelty at test time
  • Mar 26, 2026Sakana AI — Its AI Scientist research was published in Nature — reported that a paper that passed review at an ICLR 2025 workshop had an average score of 6.33, above 55% of human-written papers SourceAI-generated papers passing review entered the formal scholarly record. The debate over how to evaluate automated research got going in earnest
  • Mar 9, 2026Yann LeCun / AMI Labs — Raised $1.03 billion at a pre-money valuation of $3.5 billion SourceLLM skepticism capitalized as a $1 billion-scale bet on an alternative route
  • Feb 13, 2026Dario Amodei (Dwarkesh) — "We are near the end of the exponential" — at the same time, puts the probability of a country of geniuses in a datacenter within 10 years at 90% SourceA logical structure in which the end of scaling and the imminence of AGI are two sides of the same claim
  • Feb 12, 2026Google DeepMind — Gemini 3 Deep Think upgrade — ARC-AGI-2 84.6%, HLE 48.4% SourceARC-AGI-2 is effectively saturated. The announcement does not discuss AGI, framing the upgrade instead as accelerating science, research and engineering
  • Dec 5, 2025ARC Prize — ARC Prize 2025 results — Kaggle winner 24.03%, Claude Opus 4.5 37.6%, Poetiq refinement loop 54% SourceShowed that the key technical change of 2025 was iterative loops of search and checking
  • Nov 19, 2025Yann LeCun — Resigned as Meta's chief AI scientist and founded a startup centered on world models SourceThe biggest skeptic camp splitting off organizationally
②49 records

RAG, Fine-Tuning & Context

  • Jul 9, 2026Ashwin Gerard Colaco, Nada Lahjouji — What to Keep, What to Forget published — frames KV cache management, prompt pruning and agent memory consolidation together as a rate-distortion problem SourceOrganizes previously scattered research on context compression into a single theoretical framework
  • Jun 29, 2026Shijie Xia et al. — Diagnosing and Mitigating Context Rot in Long-horizon Search published — diagnoses agents giving up early, before the context window is used up; 7 mitigation techniques improve results by 2.6–4.9% SourceQuantifies Context Rot in the setting of search agents, and also shows that a fundamental fix is hard
  • May 12, 2026Sam Martin, Fabien Roger — Published Classifier Context Rot — frontier models missed dangerous behavior that came after 800,000 tokens of harmless activity up to 30× more often SourceEmpirical evidence that long-context limits also break down safety monitoring performance
  • Apr 19, 2026Jingbo Sun et al. — Presented AutoSearch — uses reinforcement learning to decide the minimum number of search steps based on question difficulty SourceA successor in the Search-R1 line. Made the number of searches itself something to learn
  • Mar 7, 2026Saroj Mishra et al. — Published SoK: Agentic Retrieval-Augmented Generation — formalizes the retrieval loop as a Markov decision process and catalogs vulnerabilities such as memory poisoning and retrieval misalignment SourceThe first systematic taxonomy of agentic RAG. A reference point for follow-up research
  • Jan 12, 2026Nikhil Verma — Published "Active Context Compression" — the agent consolidates knowledge on its own and prunes the original context, cutting tokens by 22.7% SourcePart of a shift in who manages context: from the pipeline to the agent itself
  • Jan 12, 2026Pietro Ferrazzi et al. — Published "Is Agentic RAG worth it?" — compares enhanced RAG and agentic RAG across many scenarios SourceEmpirical evidence that agentic RAG is not always better. A disconfirming result that grounds design choices
  • Jan 9, 2026Elias Lumer et al. — Published "Don't Break the Cache" — measured prompt caching in the APIs of 3 providers: cost savings of 41–80%, though latency can increase depending on the placement strategy SourceThe first measurement comparing the 3 providers under identical conditions since caching became a product
  • Jan 5, 2026Yi Yu et al. — Published "Agentic Memory" — exposes storing, retrieving, updating, summarizing and discarding as tool calls, and trains them with step-wise reinforcement learning SourceA case in which what to remember and what to discard is learned rather than decided by people
  • Oct 1, 2025Minki Kang et al. — ACON published — iterative compression guided by failure analysis cuts the tokens in long-horizon agents' observation and action histories by 26–54% SourceA direction that shrinks context without fine-tuning, so that even small models can be used as long-horizon agents
  • Sep 29, 2025Anthropic — Context editing and a memory tool launched in beta on the Claude API — 84% less token consumption in a 100-turn web search evaluation SourceContext management becomes a built-in API feature
  • Jul 18, 2025Manus (Yichao Ji) — Context Engineering for AI Agents published — design around the KV-cache hit rate, mask logits instead of removing tools, use the file system as context, manipulate attention through recitation, keep traces of failures in the context SourceA primary record disclosing the context-engineering principles of an agent running in real production
⑱46 records

AI Infrastructure Software

  • Aug 11, 2026Modular — Released Mojo 1.0 and, on Aug 18, 2026, open-sourced the entire compiler, tooling and standard library under Apache 2.0 SourceThe language is stabilized and fully opened 3 years after its 2023 announcement. An alternative language to CUDA C++ offers a production-grade stability contract for the first time
  • Jun 24, 2026Qualcomm / Modular — Qualcomm announced its acquisition of Modular (terms not disclosed; completed 2026-07-29) SourceThe event in which a CUDA-alternative stack went from an independent startup to an asset of a major silicon vendor
  • Jun 16, 2026AMD — Announced MLPerf Training v6.0 results — MI355X within 5% of NVIDIA B200 on Llama 2-70B fine-tuning and within 6% on Llama 3.1-8B pretraining (results submitted by AMD) SourceMaterial showing on a standard benchmark that ROCm narrowed the gap to single-digit percentages in 1 year. A real crack in CUDA's monopoly
  • Nov 30, 2025vLLM project — Announced vLLM-Omni — an omni-modal serving framework that separates encoders, the LLM core and modality generators (supports Qwen-Omni, Qwen-Image and others) SourceA sign that open-source serving infrastructure, once text-only, is expanding to cover non-autoregressive and multimodal generation
  • Nov 11, 2025NVIDIA — Released NCCL 2.28 — a Device API that lets kernels initiate communication directly, GPU-Initiated Networking, Copy Engine collectives that use no SMs, and symmetric memory support SourceA release that supports fusing communication and compute at the kernel level, changing how the communication bottleneck in large-scale distributed training and inference is tackled
  • Sep 29, 2025NVIDIA (Felix Abecassis et al.) — Pretraining Large Language Models with NVFP4 published — a 12B model trained on 10T tokens (the longest 4-bit training run in public documents), matching the FP8 baseline in training loss and accuracy SourceThe strongest positive evidence that FP4 training actually works. However, it is the vendor's own report, and no independent reproduction has been confirmed
  • Aug 6, 2025PyTorch Foundation — PyTorch 2.8 released — a limited stable libtorch ABI, control-flow operators (cond/while_loop/scan), an Inductor CUTLASS backend. 4,164 commits from 585 contributors since 2.7 SourceA release that shows the move to a compiler-centered stack, with a stabilized ABI for C++ extensions and compilable control flow
  • Jun 20, 2025PyTorch (Tristan Rice, Howard Huang et al.) — With torchft + TorchTitan on 300 GPUs, training efficiency was 81.2% with a failure injected every 60 seconds, and training still converged with failures injected every 15 seconds (1,015 failures). Weights were recovered from peers without reloading checkpoints SourceAn attempt to replace the checkpoint-based recovery paradigm. Treats each step as a distributed transaction
  • Jun 4, 2025AMD — First MLPerf Training submission — MI325X up to 8% ahead of NVIDIA H200, MI300X on par with H100 (results submitted by AMD) SourceThe first primary source making ROCm's maturity checkable on a standard benchmark. The debate over CUDA alternatives moved from claims to measurements
  • Nov 7, 2024Tanishq Kumar et al. (Harvard) — "Scaling Laws for Precision" published — proposed precision-aware scaling laws in which low-precision training reduces a model's effective parameter count. The more training data, the greater the degradation from post-training quantization SourceA study that tested, at the level of scaling laws, whether FP8/FP4 works. Redefined precision not as a free optimization but as a trade against capacity
  • Oct 9, 2024Wanchao Liang et al. (Meta) — TorchTitan presented — modular 3D parallelism and elastic scaling. An additional 30% speedup with 3D parallelism for 405B on 512 GPUs SourceThe point at which a PyTorch-native training framework became established as an alternative to Megatron-LM. FSDP2, torch.compile and FP8 were bundled into a single stack
  • Jul 31, 2024Meta (Llama 3 team) — The Llama 3 Herd of Models published — over 54 days of pretraining on 16K GPUs there were 466 job interruptions in total, 419 of them unexpected, with about 78% attributable to hardware. Even so, effective training time exceeded 90% SourceThe first primary source to publish quantitative data on fault tolerance in large-scale training. Settled, for the whole industry, that checkpointing and automatic recovery are essential
⑫43 records

Training Data & Post-Training

  • Aug 21, 2026D. Mass. (Saylor) — Motion to dismiss denied in UMG et al. v. Suno SourceThe court found the output-infringement allegations sufficient; the DMCA claims survive
  • Jul 21, 2026N.D. Cal. — Final approval of the $1.5 billion Bartz v. Anthropic settlement SourceAbout 400,000 books at $3,000 per book; more than 91% of eligible authors filed claims
  • Jul 9, 2026NYT, Daily News — Filed a motion for sanctions, alleging that OpenAI concealed evidence SourceThe dispute widens from fair use to the preservation and concealment of evidence
  • Mar 2, 2026US Supreme Court — Denied certiorari in Thaler v. Perlmutter SourceThe human authorship requirement for works generated by AI alone is settled
  • Feb 23, 2026Anthropic — Accused Chinese labs of distillation attacks — about 24,000 fraudulent accounts and more than 16 million conversations SourceThe distillation dispute moves from suspicion to quantified accusation. Choosing to disclose its detection work and seek an industry-wide response rather than sue reflects the limits of legal remedies
  • Dec 10, 2025OLG Hamburg — Dismissed the appeal in Kneschke v. LAION SourceExplicitly applied the commercial TDM exception and found that a natural-language opt-out did not meet the machine-readability requirement. Not final, as an appeal to the BGH was permitted
  • Dec 1, 2025Nous Research — Released Hermes 4.3 36B — trained the same model both centrally (FSDP+AdamW) and on Psyche (TP+DisTrO) and compared them. The Psyche run completed at an average of 144k tok/s on 24 nodes and beat the centralized run on downstream performance SourceThe first controlled comparison showing distributed training is not just possible but equal or better
  • Nov 20, 2025AI2 — Olmo 3 — released the Dolma 3 pretraining and Dolci post-training suites SourceReleased the data and the post-training pipeline together
  • Nov 11, 2025LG München I (Munich Regional Court I) — First-instance judgment in GEMA v. OpenAI SourceThe first case in Europe to treat memorization as reproduction and to rule out the TDM exception
  • Oct 27, 2025S.D.N.Y. (Stein) — In re OpenAI Copyright Infringement Litigation (MDL 25-md-3143) consolidated SourceSeven cases, including those brought by the NYT and the Authors Guild, combined into one MDL
  • Oct 2025Salesforce AI Research, CMU — Webscale-RL — automatically converts pretraining documents into verifiable QA pairs SourceMatches continual pretraining with up to 100 times fewer tokens — a direct technical answer to data exhaustion
  • Sep 29, 2025Psyche Team — Psyche shifted from proof to performance — a new trainer abstraction adds support for SFT and RL; training of Hermes 4 on ByteDance Seed-OSS-36B began on the Psyche network SourceDistributed training moves beyond pretraining demos into the post-training path of real product models
⑧41 records

Inference Infrastructure & Serving

  • Aug 24, 2026NVIDIA — Groq 3 LPX entered full production SourceA low-latency unit for agents is integrated into the rack
  • Aug 22, 2026vLLM — Introduced a large-scale sharded weight-transfer engine based on Ray Direct Transport SourceMulti-node model loading and weight swapping (RL rollouts) are promoted to first-class problems of the serving stack
  • Jul 23, 2026AMD — Launched the Helios rack-scale system and the Instinct MI455X Source72 GPUs per rack, a single scale-up domain and 31TB of HBM4 — alternative silicon entering real-world deployment
  • May 31, 2026NVIDIA — Vera Rubin NVL72 entered volume production SourceIntegrates the Vera CPU, Groq 3 LPX, BlueField-4 and Spectrum-6, with agent workloads as an explicit target
  • Apr 22, 2026Google Cloud (Cloud Next) — Previewed its 8th-generation TPUs — split into TPU 8t (training) and TPU 8i (inference) SourceSplit training and inference silicon for the first time; a signal that the era of general-purpose accelerators is ending
  • Mar 31, 2026Google Cloud — TPU7x generally available (GA) — 192GB HBM per chip, 7.37TB/s SourceIronwood went from announcement to actual supply
  • Mar 24, 2026vLLM — Model Runner V2 — a modular execution core with GPU-native input preparation and asynchronous scheduling SourceThe second rewrite of the execution layer since the V1 redesign. Serving competition moved from kernels to schedulers and execution cores
  • Mar 21, 2026Hugging Face — Archived the text-generation-inference repository SourceThe README recommends moving to vLLM, SGLang, llama.cpp or MLX; the serving engine market consolidated into two leaders plus local
  • Mar 5, 2026Zadouri, Shah, Thakkar, Dao et al. — FlashAttention-4 — co-design for asymmetric hardware scaling SourceThe latest generation of the kernel lineage, formally published at MLSys 2026
  • Dec 24, 2025NVIDIA, Groq — Non-exclusive licensing of inference technology, with key staff including Jonathan Ross brought on board SourceNon-GPU inference IP is absorbed. Groq continues as an independent company
  • Nov 6, 2025Google Cloud — Announced Ironwood (TPU v7) — 9,216 chips per superpod SourceA TPU for the inference era. Actual GA came on Mar 31, 2026
  • Oct 13, 2025NVIDIA — DGX Spark announced (shipping Oct 15) — 128 GB of unified memory SourceInference on 200B models and fine-tuning of 70B models on a desktop; local machines become agent runtimes
⑪40 records

Model Architecture & Training Methods

  • Aug 7, 2026Videau et al. — Skaling — an interaction exponent bridging Chinchilla's exponents and Kaplan's coupling SourceBreaks the assumption that N and D are independent, reducing prediction error by 1.5 to 3 times and extrapolating with about 10 times less compute
  • Jul 31, 2026Google DeepMind — DiffusionGemma — an open-weight diffusion language model fine-tuned from the MoE Gemma 4 SourceShows a path to obtaining a diffusion model with less than 10% of the original token budget; about 1,500 tok/s on a single H100
  • May 27, 2026Mistral AI — Brought in physics AI by acquiring and integrating Emmi AI — learns from physics solver outputs to predict physical behavior directly from geometry and boundary conditions SourceAn LLM lab expanding into replacing industrial simulation. No named model release at the time of the announcement
  • Apr 26, 2026DeepSeek-AI — DeepSeek-V4 — CSA and HCA hybrid attention, mHC, Muon, 32T+ tokens SourceAt 27% of the per-token FLOPs and 10% of the KV cache of V3.2, it made 1M context something supported all the time
  • Mar 16, 2026Mistral AI — Released Mistral Small 4 — 119B total / 6B active MoE, 256k context, Apache 2.0. Merges Magistral (reasoning), Devstral (coding) and Small (instruct) into one, with reasoning effort adjustable as a parameter SourceBrought the trend of merging separate model families into a single set of weights to open weights
  • Dec 31, 2025Google (Behrouz et al.) — Nested Learning / HOPE SourceRecasts the optimizer as an associative memory module and proposes a sequence model that learns a self-update algorithm
  • Dec 4, 2025Google Research — Introduced Titans and MIRAS, a theoretical framework that generalizes it — a neural long-term memory architecture SourceThe route of working around long-context limits through architecture rather than retrieval
  • Oct 30, 2025Moonshot AI — Kimi Linear — a layer-wise hybrid of KDA and MLA SourceThe first to beat full attention in a fair comparison; up to 75% less KV cache and up to 6× faster decoding at 1M context
  • Sep 19, 2025xAI — Grok 4 Fast released — reasoning and non-reasoning modes unified in one set of weights and switched via the system prompt. 2M-token context, 40% fewer thinking tokens SourceAn early case of merging reasoning and non-reasoning models into a single set of weights — a direction that later became the industry standard
  • Jul 28, 2025Moonshot AI — Kimi K2 — a 1T MoE trained on 15.5T tokens with MuonClip, without a single loss spike SourceMuon validated at 1T scale
  • Jun 17, 2025Inception Labs et al. — Mercury announced — with diffusion-based parallel token generation, Mercury Coder Mini reaches 1,109 tok/s on an H100 SourceClaims up to 10× the speed of speed-optimized frontier models. The first time a diffusion LM entered a commercial track
  • Feb 20, 2025Naver — Released a new HyperCLOVA X model — parameters cut to about 40%, operating costs improved by more than 50%, MMLU 79.6%, video understanding added SourceA turning point toward efficiency for a Korean lab
⑯38 records

Evaluation & Benchmarks

  • Aug 27, 2026Surge AI — Published "LMArena is a cancer on AI" — reporting disagreement with 52% of 500 votes in its own sample, and strong disagreement with 39% SourceA critique that leaderboards are optimized for formatting and impressions rather than accuracy. But this is the company's own analysis as a seller of human data, not an independent check
  • Jul 23, 2026deedy (imo-2026 community project) — Ran the IMO 2026 problems autonomously on 7 frontier models and published the audit trails and independent grading — Claude Fable 5, GPT-5.6 Sol (xhigh) and Kimi K3 were checked at a perfect 42/42 (runs Jul 17–23) SourceA community-led release of audit trails, rather than a lab announcement, plus real-time olympiad evaluation with independent grading emerges as a new practice for checking results
  • Jul 6, 2026Edwin H. Wintermute et al. — BioSecBench-Refusal released — a benchmark that measures, as a pair, the ability to identify hazards in biology research tasks and the rate of over-refusal on legitimate dual-use research SourceTies capability and refusal behavior into a single paired metric, formally bringing the over-refusal problem into biosecurity evaluation
  • Jun 12, 2026Epoch AI — Released a revised FrontierMath — in Tier 4, 12 problems revised and 7 removed (in Tiers 1–3, 123 problems revised and 5 removed) SourceLarge-scale errors were found even in a top-difficulty benchmark, raising doubts about the reliability of its review once again
  • Jun 9, 2026Andrew Bo Liu et al. — Released ABC-Bench — a benchmark that evaluates agentic LLMs' ability to carry out biology experiments (writing code for liquid-handling robots, designing DNA fragments for in vitro assembly, evading DNA synthesis screening) SourceThe arrival of a new family of benchmarks in which bio-risk evaluation moves from knowledge Q&A to evaluating an agent's ability to run an actual experimental loop
  • May 6, 2026Laude Institute — Released Terminal-Bench 2.1 — fixed 28 tasks because of changes in external dependencies, resource mismatches and specification errors SourceA case of a benchmark publicly admitting its own defects and introducing an ongoing checking process
  • Mar 25, 2026ARC Prize — Released ARC-AGI-3 — humans 100% versus frontier AI 0.51% SourceThe axis moved from static problem solving to interactive exploration and skill acquisition without instructions
  • Feb 23, 2026OpenAI Frontier Evals — Why SWE-bench Verified no longer measures frontier coding capabilities SourceAt least 59.4% of the audited problems had flawed tests; the vendor acknowledges that the benchmark harness's signal is contaminated
  • Jan 29, 2026METR — Released Time Horizon 1.1 — tasks increased from 170 to 228; evaluation infrastructure switched from Vivaria to Inspect. Only 14 of the previous 33 models were re-evaluated. The doubling time for the period since 2024 went from 109 days to 89 days SourceA case of the issuing organization itself showing that a change in methodology alone moves the numbers. The top model's 50% time horizon of about 5.3 hours is based on this revision and cannot be compared directly with the previous version
  • Jan 28, 2026LMArena — Renamed LMArena to Arena SourceOfficially broadens its identity from a text LLM arena to a multi-modality evaluation platform
  • Jan 22, 2026METR — Published an official note setting out the limitations of the time horizon metric — states an error range of about 2x and that the task distribution is unclear SourceA case of the issuing organization itself putting the brakes on its own metric being over-generalized when cited
  • Nov 7, 2025Terminal-Bench team — Released Terminal-Bench 2.0 and Harbor, an evaluation harness SourceEvaluation infrastructure that adds cloud container execution and an RL/SFT rollout interface to agent CLI evaluation
③37 records

Tools & Agents

  • Aug 25, 2026Ollama, Anthropic — Claude Desktop added support for Ollama as a third-party gateway provider SourceA closed client accepts a local open-model runtime as an official backend
  • Jul 30, 2026OpenClaw (Kevin Lin) — Introduced an extended-stable channel (security and stability backports) and a public Maturity Scorecard SourceAn agent project that had kept a rapid-release culture moves to a system of LTS releases and maturity disclosure
  • Jul 13, 2026OpenClaw — Release 2026.7.1 — a ClawRouter routing plugin, a session-first redesign of the Control UI, and entry into control-plane-safe mode after repeated abnormal boots SourceRecovery safeguards on the level of consumer software come to an agent runtime
  • Jul 8, 2026OpenClaw Foundation — Launched as a 501(c)(3) nonprofit foundation (chair: Dave Morin). Steinberger keeps technical decision-making authority. Donors include OpenAI and the University of Michigan, among others SourceA structure in which competing labs jointly sponsor a neutral agent standard
  • Jun 3, 2026OpenAI — Announced the end of AgentKit's Agent Builder and Evals products and pointed users to the Agents SDK and Workspace Agents SourceProducts announced in October 2025 were pulled after 8 months — the instability of the agent tooling layer
  • Feb 14, 2026Peter Steinberger — The creator announced he was joining OpenAI and transferred the project to the OpenClaw Foundation SourceThe transition from a personal project to foundation governance happened at the same time as a hire by a frontier lab
  • Jan 29, 2026OpenClaw — Renamed again, from Moltbot to OpenClaw, and launched an official blog SourceThe 3rd renaming in 2 months. A turning point that separates the project's identity from a personal brand
  • Jan 27, 2026Peter Steinberger — Renamed Clawdbot to Moltbot following a trademark objection from Anthropic SourceA case where trademark law put a stop to the practice of open-source agents riding on a model provider's brand
  • Nov 4, 2025Anthropic — Published Code Execution with MCP — exposes MCP servers as code APIs so only the needed tool definitions are loaded. In the example, 150,000 → 2,000 tokens SourceA structural fix for tool definitions eating up the context
  • Nov 2025Peter Steinberger — Released Warelay, a personal AI assistant, which later led to the Clawd → Clawdbot lineage SourceThe starting point that popularized the form of a locally resident agent that uses messaging apps as its interface
  • Oct 22, 2025LangChain — LangChain 1.0 and LangGraph 1.0 released — create_agent as the standard loop, middleware that hooks into the agent loop, durable state SourceFormalizes customization of the agent loop as the concept of middleware
  • Sep 11, 2025Anthropic (Ken Aizawa) — Writing Effective Tools for Agents published — choosing which tools to build, namespacing, returning meaningful context, token efficiency, and prompt-engineering tool descriptions SourceA document showing that prompt engineering has moved into the design of tool definitions
⑦37 records

Multimodal

  • Aug 7, 2026xAI — Imagine Image 2.0 released — multi-reference editing (up to 5 images), segmentation, background removal SourceAlthough it is an image model launched after regulatory pressure over NCII, the announcement makes no mention of watermarking or moderation
  • Aug 2, 2026EU (AI Act Article 50) — The originally scheduled application date of the obligations for machine-readable marking of AI-generated content and for deepfake disclosure. In practice, it was postponed to Dec 2, 2026 by Omnibus Regulation (EU) 2026/1744 SourceProvenance labeling such as C2PA and SynthID moves from goodwill to a regulatory compliance requirement. But citing Aug 2, 2026 as the date of application is wrong
  • May 7, 2026OpenAI — GPT-Realtime-2 / -Translate / -Whisper SourceBrought reasoning effort and parallel tool calls to voice models and raised context from 32K to 128K. Voice became an environment for running agents
  • Apr 23, 2026OpenAI — GPT-5.5 — OSWorld-Verified 78.7%, MMMU-Pro 81.2% SourceShowed in numbers that visual understanding and operating a computer are the same capability
  • Apr 17, 2026Alibaba Qwen — Qwen3.5-Omni technical report — Thinker-Talker Hybrid-Attention MoE, 256k context SourceOpen-weight omni-modal models caught up with the frontier, handling 10 hours of audio and 400 seconds of video
  • Apr 16, 2026Anthropic — Claude Opus 4.7 — high-resolution vision up to 2,576px on the long edge SourceShowed that reading dense screenshots and diagrams had been the bottleneck in computer use performance
  • Jan 29, 2026Moonshot AI — Released Kimi K2.5 — MoE with 1T total / 32B active parameters, natively multimodal through pretraining on vision-language tokens, introduces a self-coordinating Agent Swarm, modified-MIT SourceThe point where the K2 line shifted from text to native multimodality, and the arrival of an open-weight agent swarm
  • Jan 28, 2026xAI — Released the Grok Imagine API — text-to-video, image-to-video, and editing features such as object manipulation and restyling SourceCommercializes a video generation API in the midst of regulatory pressure over NCII. The announcement makes no mention of content safety or moderation
  • Nov 25, 2025Black Forest Labs — FLUX.2 — dev 32B with open weights SourceCombines a Mistral-3 24B VLM with a rectified flow transformer; a release in which open weights narrowed the quality gap
  • Nov 20, 2025Google — Nano Banana Pro (Gemini 3 Pro Image) SourceAccurate infographics and multilingual text rendering redefined image generation as a reasoning task; ships with SynthID
  • Nov 18, 2025Google — Gemini 3 Pro — MMMU-Pro 81%, Video-MMMU 87.6% SourceA natively multimodal design paid off as an edge in video and document understanding
  • Nov 12, 2025World Labs — Launched Marble to the public SourceGenerates editable 3D worlds from text, images and video, and exports them as Gaussian splats and meshes
④31 records

MCP & Skills Standardization

  • Aug 17, 2026Agentic AI Foundation, Google — Google's Agent2Agent (A2A) protocol was transferred to the Agentic AI Foundation under the Linux Foundation (host of MCP, AGENTS.md and goose; 250+ members) (Axios exclusive) SourceWith MCP and A2A unified under a single neutral foundation, the governance of agent standards is reorganized around a foundation
  • Jul 28, 2026MCP — Specification 2026-07-28 — making the protocol core stateless SourceRemoves initialize and Mcp-Session-Id, introduces MRTR, adds server/discover, and promotes the Tasks and MCP Apps extensions
  • Jul 28, 2026Model Context Protocol — Specification revision — removes protocol-level sessions and Mcp-Session-Id, drops the initialize handshake, makes server/discover mandatory, has the MRTR pattern replace server-initiated requests, announces the end of support for Roots, Sampling, Logging and OAuth DCR, and introduces a 12-month deprecation grace policy SourceMCP is redesigned from the ground up, from a protocol of stateful sessions into a stateless, per-request one. The largest break in backward compatibility yet
  • Jun 1, 2026NVIDIA / OpenClaw (Vincent Koc) — Integrated Skill Card (an open trust artifact carrying the publisher's identity, capabilities and scan results) and SkillSpector (detects hidden instructions and capability mismatches using static analysis + LLM semantic analysis) into ClawScan SourceA skill distribution standard was proposed as a vendor-led open spec, beyond a single project
  • Apr 2, 2026AAIF, Linux Foundation — MCP Dev Summit North America 2026 (New York, 1,200 attendees) SourceAAIF reached 170 member companies in 4 months, MCP SDK monthly downloads passed 110 million, and security — including a DNS rebinding demo — was the biggest topic
  • Jan 26, 2026Anthropic, MCP — Claude fully supports MCP Apps as an official extension SourceA protocol that used to carry only text and structured data now extends to on-screen interfaces
  • Dec 18, 2025Anthropic — Published Agent Skills as an open standard (agentskills.io) SourceReleased a specification and a reference SDK; the second standardization attempt after MCP
  • Dec 9, 2025Anthropic, Block, OpenAI — Donated MCP to the AAIF under the Linux Foundation SourceThe founding projects were MCP, goose and AGENTS.md, at a point when public MCP servers exceeded 10,000
  • Nov 25, 2025Model Context Protocol — Spec revision — OIDC Discovery support, incremental scope consent via WWW-Authenticate, OAuth Client ID Metadata Documents recommended, URL mode elicitation, experimental tasks, formalized governance and working groups SourceAligned the auth model with standard OAuth practice and moved spec governance to committee procedures
  • Nov 24, 2025Anthropic — advanced tool use beta — Tool Search Tool, Programmatic Tool Calling SourceCut tool tokens by 85% while MCP evaluation accuracy rose from 49% to 74% on Opus 4 — the paradox that showing fewer tools makes the model more accurate
  • Nov 21, 2025Anthropic, OpenAI, MCP-UI (jointly) — Proposed the MCP Apps extension (SEP-1865) SourceLets servers provide interactive UIs via the ui:// scheme and sandboxed iframes; co-authored by competing camps
  • Oct 16, 2025Anthropic — Agent Skills launched (SKILL.md, progressive disclosure) SourcePackages procedural knowledge as folders and loads it only when needed, in three stages: metadata, body, bundled files
⑤24 records

Loop & Harness Engineering

  • Aug 19, 2026OpenAI (Bonamy, Choi) — Codex as a platform: build on the open agent harness SourceWith the Codex CLI, app-server, SDK and MCP open-sourced, the harness is established as a platform layer
  • Jul 14, 2026Latent Space (Richard MacManus) — AIE World's Fair 2026 trends — from agents to the systems around agents SourceThe inner loop vs. outer loop framing is elevated to the industry's standard narrative
  • Jun 26, 2026Andrew Ng (The Batch 359) — A 3-layer loop — agentic loop (minutes), developer feedback loop (hours), external feedback loop (days) SourceEstablishes the popular definition of loop engineering
  • Jun 23, 2026Armin Ronacher — The Coming Loop SourceThe representative critique that loops work for checkable, short-lived tasks but are fragile for code that has to be maintained
  • Jun 17, 2026Databricks — What is an AI Agent Harness? — defines the harness in 8 components SourceTakes an enterprise view, counting observability and governance as part of the harness
  • Jun 8, 2026de Macedo (arXiv:2606.10106) — What makes a harness a harness — T1–T4 necessary and sufficient conditions SourceProblematized the term's many meanings academically, proposing as the condition a control mechanism that does not depend on the model's compliance
  • Jun 7, 2026Addy Osmani — Loop Engineering — 5 components plus State and Memory SourceEstablished the term loop engineering; a shift in role from prompt writer to loop designer
  • May 15, 2026Addy Osmani (O'Reilly Radar) — Agent Harness Engineering SourceThe containment position: the harness is the delivery vehicle of context engineering
  • May 7, 2026Zhang et al. (arXiv:2605.23950) — Stop Comparing LLM Agents Without Disclosing the Harness SourceHarness-induced variance was 7.80× model-induced variance, and rankings flipped in 6 of 9 model pairs
  • Apr 2, 2026Birgitta Böckeler (martinfowler.com) — Harness engineering for coding agent users — Agent = Model + Harness SourceThe standard reference for the user-side definition of the harness as everything in an AI agent except the model
  • Mar 24, 2026Anthropic Labs (Prithvi Rajasekaran) — Harness design for long-running application development Source3 agents — Planner, Generator, Evaluator; a single agent ran 20 minutes for $9 versus 6 hours for $200 with the full harness
  • Mar 12, 2026HumanLayer — Skill Issue: Harness Engineering for Coding Agents SourceDefined harness engineering as a subset of context engineering, with the concepts of a context firewall and back-pressure
⑲18 records

On-Device & Local Inference

  • Jul 9, 2026Ollama — Raised $89 million; 8.9 million developers SourceThe local-inference deployment layer stands on its own as independent commercial infrastructure
  • Jun 9, 2026Apple — Announced an expansion of the Foundation Models framework at WWDC 2026 — image input, integration of server-side third-party models (Claude, Gemini) through the same Swift API, free Private Cloud Compute for small developers, and a plan to open-source the framework (as announced) SourceOn-device AI frameworks shifted from a closed single model to a hybrid, multi-provider platform
  • Jun 5, 2026Ollama — In 0.30, added llama.cpp-based GGUF compatibility to its MLX engine SourceThe local deployment stack reorganized from sole reliance on llama.cpp to MLX first + GGUF compatibility
  • May 29, 2026llama.cpp community — Discussion #23853 — released a tool for NVFP4/MXFP6 mixed precision + imatrix per-layer optimization. Comparisons shared in the discussion found NVFP4 faster but below Q4_K_M in quality SourceReal measurements showing hardware-native FP4 has not yet beaten community quantization. The format race is unresolved
  • Apr 9, 2026Microsoft — Foundry Local generally available (GA) — a cross-platform on-device inference stack for Windows, Linux and macOS, with automatic GPU, NPU and CPU acceleration and native Metal on Apple Silicon SourceMade local inference, free of cloud lock-in and per-token billing, an official product line
  • Apr 2, 2026Google (Android) — Released Gemma 4 as an AICore developer preview — E4B/E2B, more than 140 languages, multimodal; claims up to 4× faster than the previous generation and up to 60% battery savings (preview, not GA) SourceOn-device AI on Android evolved into a system in which a system service (AICore) distributes open models directly
  • Mar 11, 2026llama.cpp (ggml-org) — Merged PR #19769 — introduced GGML_TYPE_NVFP4. FP4 E2M1 weights + UE4M3 block scales, loading NVIDIA ModelOpt outputs directly SourceGGUF absorbed a hardware-native FP4 format beyond its own K-quants
  • Jan 11, 2026Uygar Kurt — Published a paper evaluating llama.cpp quantization in a unified way — compares 3- to 8-bit K-quants and legacy formats across benchmarks, perplexity, throughput, file size and even quantization time SourceThe first unified reference to move the choice of quantization from gut feel to measurement
  • Nov 19, 2025Apple — Published M5 GPU Neural Accelerator + MLX benchmarks — time to first token (TTFT) for LLM inference up to 4× faster than M4, memory bandwidth 153GB/s SourceMaterial in which a consumer silicon generation change began to be promoted directly by local LLM inference performance
  • Oct 19, 2025llama.cpp community — NVFP4 support requested in Discussion #16668. A maintainer pointed to the MXFP4 support PR as the template SourceThe point at which the path was set for GGUF to absorb vendor FP4 formats one after another
  • Sep 19, 2025Ollama — Cloud models preview — keeps the local tool chain while offloading only large models to data-center hardware SourceA local deployment tool expands into hybrid execution. The point at which the definition of running locally began to blur
  • Jun 26, 2025Google — Gemma 3n officially released — E2B runs in about 2 GB and E4B in about 3 GB of memory; takes image, audio, video and text input; E4B is the first model under 10 billion parameters to pass 1300 points on LMArena SourceA milestone at which small, on-device-class models began to combine multimodality with competitive benchmark scores
㉑14 records

Graph Engineering

  • Jun 24, 2026Chen, Razkenari et al. (AWS-affiliated) — Is GraphRAG Needed? published — systematically analyzes the real gains of graph and agentic RAG over basic RAG, and reports that expanding retrieval does not carry over proportionally into generation quality SourceBrings into the mainstream the self-criticism that evaluations centered on retrieval metrics have overstated the benefits of graph techniques
  • Apr 1, 2026Dongzhe Fan et al. — Published Do We Still Need GraphRAG? — used the RAGSearch benchmark to compare dense RAG and GraphRAG, each combined with agentic search SourceShowed that agentic search narrows the gap between the two approaches, updating the 2024 GraphRAG debate
  • Jun 6, 2025Xiang, Wu et al. (Hong Kong Polytechnic University and others) — Published the paper When to use Graphs in RAG and released GraphRAG-Bench — systematically identified the conditions under which GraphRAG falls behind vanilla RAG on many real tasks SourceThe turning point into a phase of testing with benchmarks when graphs actually help — the methodological forerunner of the 2026 skepticism
  • Apr 28, 2025Mem0 (Chhikara, Khant et al.) — Published the Mem0 paper — alongside the agent long-term memory system, proposed a graph-based variant (Mem0g) that captures the relational structure among conversation elements (performance advantage is company-reported) SourceA commercial agent long-term memory product adopted graph representations as a core extension — graph engineering spreading beyond RAG into the memory layer
  • Nov 25, 2024Microsoft Research (Edge, Trinh, Larson) — LazyGraphRAG released — combines best-first and breadth-first search without upfront summarization; claimed its indexing cost is 0.1% of full GraphRAG and that it cuts global query cost by more than 700 times SourceThe original GraphRAG authors acknowledged the cost problem themselves and turned toward deferred graph construction
  • Oct 8, 2024HKU (Guo, Xia, Huang et al.) — LightRAG released — a dual-level retrieval and incremental-update framework combining graph structures with vector representations SourceThe open-source camp's leading lightweight alternative to the costly indexing of Microsoft GraphRAG
  • Oct 1, 2024gusye1234 (open-source community) — Release of nano-graphrag v0.0.8, which reimplemented GraphRAG's core features in about 1,100 lines of code SourceCommunity complaints that the official implementation was heavyweight surfaced as a hackable, ultra-light reimplementation
  • Sep 10, 2024Ant Group (Liang, Sun et al.) — KAG (Knowledge Augmented Generation) released — a framework that combines knowledge graphs and vector retrieval to raise LLM performance in professional domains SourceA case of Chinese Big Tech industrializing the GraphRAG lineage for professional domains (finance, healthcare)
  • Aug 15, 2024Peng, Zhu et al. (Peking University, Ant Group, and others) — Graph Retrieval-Augmented Generation: A Survey published — the first comprehensive GraphRAG survey, organizing the graph-based indexing, retrieval and generation workflow SourceBrought a sprawl of variants under one taxonomy, formalizing graph-based RAG as a research field of its own
  • Jul 2, 2024Microsoft Research — Open-sourced GraphRAG on GitHub and provided a solution accelerator for deployment on Azure SourceThe point at which a research technique spread as a widely replicable implementation
  • May 23, 2024Bernal Jiménez Gutiérrez et al. (Ohio State University, Stanford) — HippoRAG published — integrates multiple hops in a single retrieval step using a knowledge graph and personalized PageRank SourceThe multi-hop retrieval branch of the GraphRAG lineage
  • Apr 24, 2024Darren Edge et al. (Microsoft Research) — GraphRAG paper published — detects communities in an entity graph and combines community summaries via map-reduce to answer global queries SourceHandles global summarization queries that local retrieval cannot answer

Showing 264 here; 790 are not listed. See them all in all records from that week.

How this relates to the main moduleThis page holds only what was added that week. For the whole story, open the main module — it is one latest edition, revised every week, while weekly editions stay as they were.