Checked this week: 71 entries, 16 axes. Most: ⑨ Open-Weight Lineage, 27 entries. More than the past 4-week average of 22 (4 weeks: 11, 33, 26, 18). 12 new actors — OpenAI (Radford et al.), Google (Devlin et al.), Alibaba Cloud.
71records added
16axes that moved
Oct 5–11dates
1,251total 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.
⑨27 records
Open-Weight Lineage
Oct 2, 2026Allen Institute for AI (Ai2) — AstaBrief 8B released — an open-weight model for generating scientific reports, based on Qwen3-8B, released together with its training data and checkpoints as a Hugging Face collection, plus an example workflow for building reports from local PDFs. Trained on 47K supervised examples drawn from real user queries and about 6K DPO pairs. Says its answer accuracy, citation accuracy and citation recall on SQABench-CS2 (200 computer science questions) are comparable to commercial alternatives. Fast mode averages 51.1 seconds per report, about 3.5 times faster than Thinking mode at 178.5 seconds. Integrated into the Asta platform. SourceThe point at which deep-research-style capabilities came down to an 8B open-weight model. Using citation accuracy and recall as evaluation axes matters, because in report generation the problem is not sentence quality but whether the citations are real. Releasing even the training data creates a reproducible basis for comparison.
Dec 1, 2025DeepSeek — Released DeepSeek-V3.2 and V3.2-Speciale — the official successor to V3.2-Exp, with thinking during tool use; weights for both under MIT (Speciale offered on a temporary API endpoint until December 15) SourceOpen-weight reasoning models; DeepSeek claimed V3.2-Speciale reached gold-level results at IMO, IOI and other 2025 competitions
Nov 6, 2025Moonshot AI — Released Kimi K2 Thinking — 1T total / 32B active MoE, 256K context, native INT4 inference via quantization-aware training in post-training, under a Modified MIT License SourceAn open-weight reasoning agent trained to interleave thinking with tool calls; Moonshot claims it sustains 200–300 sequential tool calls without human intervention
Aug 21, 2025DeepSeek — Released DeepSeek-V3.1 — one model switching between Think and Non-Think modes, 671B total / 37B active, 128K context, 840B tokens of long-context continued pretraining on top of V3, MIT SourceBilled by DeepSeek as its first step toward the agent era; the API's deepseek-chat and deepseek-reasoner became the non-thinking and thinking modes of one model, merging what had been separate V3 and R1 offerings
Jul 24, 2025SK Telecom — Open-sourced A.X 3.1 — a 34B standard model developed from scratch, with 32K context (131K with YaRN), under Apache 2.0 SourceUnlike A.X 4.0, which continued pretraining open-source models on Korean data, it was designed and trained entirely in-house; SKT's open models now numbered four, two A.X 3.1 and two A.X 4.0
Jul 22, 2025Alibaba Qwen — Released Qwen3-Coder — first model Qwen3-Coder-480B-A35B-Instruct (480B total / 35B active MoE), 256K native context extendable to 1M, pretrained on 7.5T tokens (70% code), Apache 2.0 SourceAlibaba claimed state-of-the-art results among open models in agentic coding, browser use and tool use, and open-sourced the command-line coding tool Qwen Code alongside it
May 23, 2025Kakao — Open-sourced Kanana 1.5 — four models (base and instruct at 8B and 2.1B) under Apache 2.0 SourceThe first Kanana release under a license that allows commercial use (Apache 2.0)
Mar 18, 2025LG AI Research — Released EXAONE Deep — reasoning-specialized 32B, 7.8B and 2.4B models with 32K context, under the EXAONE AI Model License Agreement 1.1 - NC SourceLG's open-weight line specialized for math, science and coding reasoning; company-reported scores for the 32B were 94.5 on the 2025 CSAT math section and 90.0 on AIME 2024
Feb 27, 2025Kakao — Published the Kanana technical report and open-sourced Kanana Nano 2.1B (base, instruct, embedding) for research use under CC BY-NC 4.0 SourceThe first open weights in the Kanana line, released alongside a report on training the 2.1B–32.5B Kanana lineup
Jan 30, 2025Mistral AI — Released Mistral Small 3 — a latency-optimized 24B model, with both pretrained and instruction-tuned checkpoints under Apache 2.0 SourceMistral recommitted to Apache 2.0 for its general-purpose models while moving away from MRL-licensed ones; company-reported over 81% on MMLU at 150 tokens/s
Dec 9, 2024LG AI Research — Released EXAONE 3.5 as open source — three sizes, 2.4B (on-device), 7.8B and 32B, with 32K context, under the EXAONE AI Model License Agreement 1.1 - NC SourceA follow-up four months after the EXAONE 3.0 7.8B release that expanded to three sizes while keeping a non-commercial license
Dec 6, 2024Meta — Released Llama 3.3 70B — a text-only 70B model with 128K context, pretrained on about 15T tokens, supporting 8 languages, under the Llama 3.3 Community License SourceAn update to the Llama 3 line that shipped a single 70B instruction-tuned model, with no separate pretrained release
Sep 25, 2024Meta — Released Llama 3.2 — 11B and 90B vision LLMs plus lightweight 1B and 3B text models (128K context) for edge and mobile devices, under the Llama 3.2 Community License SourceThe 11B and 90B were the first Llama models to support image input, released alongside 1B and 3B models aimed at on-device use
Sep 19, 2024Alibaba Qwen — Released Qwen2.5 — seven sizes from 0.5B to 72B, pretrained on up to 18T tokens, 128K context, 29+ languages; Qwen2.5-Coder and Qwen2.5-Math released alongside SourceAll open models except the 3B and 72B were released under Apache 2.0, and the 14B and 32B sizes missing from the previous generation returned
Sep 11, 2024Upstage — Released Solar Pro Preview — 22B, designed to run on a single GPU, English-only with a 4K context window, open source with commercial use allowed (MIT) SourceA preview released as weights plus free API access ahead of the full release planned for November; Upstage claimed it competes with multi-GPU models such as Llama 3.1 70B
Jul 24, 2024Mistral AI — Released Mistral Large 2 — 123B, 128K context, dozens of languages and 80+ coding languages; instruct weights published under the Mistral Research License SourcePublished the weights of a 123B flagship designed for single-node inference, limited to research and non-commercial use, with a commercial license required for self-deployment
Jun 27, 2024Google — Released Gemma 2 (9B and 27B) under the Gemma license; the 27B runs full-precision inference on a single NVIDIA H100 or A100 80GB GPU or a TPU host SourceThe second Gemma generation; Google claimed the 27B competes with models more than twice its size and the 9B outperforms Llama 3 8B
Jun 17, 2024DeepSeek — Released DeepSeek-Coder-V2 — 236B (21B active) and 16B (2.4B active) MoE models, further pretrained from a DeepSeek-V2 intermediate checkpoint on 6T additional tokens; programming languages expanded from 86 to 338 and context from 16K to 128K SourceAn open-weight code model that DeepSeek claimed matched GPT-4 Turbo on code-specific tasks
Jun 7, 2024Alibaba Qwen — Released Qwen2 — five sizes (0.5B, 1.5B, 7B, 57B-A14B, 72B), trained on 27 additional languages beyond English and Chinese, with up to 128K context for the 7B and 72B Instruct models SourceOnly the 72B kept the original Qianwen License while the other four moved to Apache 2.0, the release where Qwen's licensing began to open up
May 29, 2024Mistral AI — Released Codestral — a 22B code model covering 80+ programming languages, with 32K context and fill-in-the-middle, under the new Mistral AI Non-Production License SourceIntroduced a non-production license that publishes weights for research and testing only, with commercial use requiring a separate license
Apr 17, 2024Mistral AI — Released Mixtral 8x22B — a sparse MoE with 141B total / 39B active parameters, 64K context and native function calling, under Apache 2.0 SourceMistral's larger MoE follow-up to Mixtral 8x7B, again under Apache 2.0
Feb 26, 2024Mistral AI — Launched Mistral Large — a flagship text model with 32K context, native fluency in English, French, Spanish, German and Italian, and function calling; offered via API on la Plateforme and Microsoft Azure, with no public weights SourceAn API flagship launched with Microsoft Azure as Mistral's first distribution partner; Mistral claimed it ranked second only to GPT-4 among models available through an API
Dec 14, 2023Upstage — Released SOLAR 10.7B — built by depth up-scaling (DUS) an open-source 7B model and pretrained on over 3T tokens; base model under Apache 2.0, instruction-tuned model under CC BY-NC 4.0 SourceA model from a Korean startup that, per Upstage, topped the Hugging Face Open LLM Leaderboard at release; the 74.2 average belonged to the instruction-tuned SOLAR-10.7B-Instruct, while the base model scored 66.04
Nov 29, 2023DeepSeek — Released DeepSeek LLM 7B and 67B (Base and Chat) — trained from scratch on 2T tokens of English and Chinese, with intermediate training checkpoints also released; commercial use permitted under the model license SourceA general-purpose DeepSeek model released down to its intermediate checkpoints; DeepSeek claimed the 67B Base outperformed Llama 2 70B Base in reasoning, coding, math and Chinese comprehension
Aug 24, 2023Meta — Released Code Llama (7B/13B/34B) — code-specialized models built on Llama 2, including Python and Instruct variants, under the same community license as Llama 2 SourceA code-specialized line trained on 500B tokens of code and code-related data on top of Llama 2; a 70B version was added on January 29, 2024
Aug 3, 2023Alibaba Cloud — Released Qwen-7B and Qwen-7B-Chat — the first open weights of the Tongyi Qianwen family, distributed on ModelScope and Hugging Face under the Tongyi Qianwen license (commercial use by application) SourceAlibaba's first open-weight language model; the original release was trained on 2.2T tokens, raised to 2.4T in an update on September 25, 2023
May 25, 2021Naver — Unveiled HyperCLOVA at NAVER AI NOW — 204 billion parameters by the company's account, with 97% of its training data in Korean SourcePresented by Naver as Korea's first hyperscale AI and a means to AI sovereignty; a closed model later succeeded by HyperCLOVA X
⑥17 records
Reasoning Models & the AGI Debate
Sep 29, 2026OpenAI — GPT-6.1 Sol released — the company claims it beats GPT-6 Sol's best scores by 6.4 percentage points on the DeepSWE v1.1 coding benchmark, by 4.8 points on AutomationBench and by 7 points on OSWorld 2.0 computer use, and scores more than 2 times higher on Terminal-Bench Science. Says it cut the factual error rate at low reasoning effort from 11.4% to 7.7%. Priced per 1M tokens at $2.00 input, $0.10 cached input and $10.00 output, it is positioned as offering intelligence close to Astra at 1/5 of GPT-6 Astra's standard rate. Available the same day across Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex, and in the API (gpt-6.1-sol). An Ultrafast version with up to 8 times faster token generation is planned for later. SourceA mid-tier update released alongside Anthropic's Sonnet 5.5 in the same week. That two labs simultaneously touted "nearly the same work at 1/5 of the frontier price" shows the competitive axis of the second half of 2026 moving from top performance to price per unit of performance.
Sep 28, 2026Anthropic — Claude Sonnet 5.5 released — the company claims 70.6% on Terminal-Bench 4.0 (Sonnet 5: 10.3%), 55.5% on CursorBench 4.0 (34.1%) and 1844 points on GDPval-AA v2.1 (1449), 2 points below Opus 5.5 on GDPval-AA. Says it is more than 30% faster than Sonnet 5 and up to 30% cheaper per task. Per 1M tokens: $2 input, $10 output, $0.20 for cache reads and $2.50 for cache writes. Available immediately on all platforms including AWS, Google Cloud and Azure (model ID claude-sonnet-5-5), with a no-data-retention option. Adds that it is the first Sonnet to beat Pokémon Red from screenshots alone. SourceThe vendor's claim itself — that the mid tier has climbed to within two points of the top tier — is this week's event. As the performance gap between tiers narrows, price becomes the deciding factor, and the reason to use frontier models shifts from task difficulty to latency and cost. But jumps such as 10.3%→70.6% on Terminal-Bench have no outside check of whether they compare like with like, so they are kept as company claims only.
Feb 17, 2026Anthropic — Claude Sonnet 4.6 released — 1M-token context window in beta, pricing unchanged at $3/$15 per 1M tokens SourceBecame the default model for Free and Pro plans in claude.ai and Claude Cowork; Anthropic reported that early-access developers strongly preferred it to its predecessor and often to Opus 4.5 from November 2025
Dec 11, 2025OpenAI — GPT-5.2 released — Instant, Thinking and Pro, pitched as a frontier model for professional knowledge work and long-running agents SourceCompany-reported for GPT-5.2 Thinking: GDPval wins or ties 70.9% (GPT-5: 38.8%), ARC-AGI-2 (Verified) 52.9% (GPT-5.1 Thinking: 17.6%), SWE-bench Verified 80.0%
Nov 24, 2025Anthropic — Claude Opus 4.5 released — $5/$25 per 1M tokens (Opus 4 and 4.1: $15/$75), with a new effort parameter in the API SourceCompany-reported: at medium effort it matches Sonnet 4.5’s best SWE-bench Verified score while using 76% fewer output tokens
Nov 12, 2025OpenAI — GPT-5.1 released — the GPT-5 series updated with two models, GPT-5.1 Instant and GPT-5.1 Thinking SourceOpenAI said Thinking adapts its thinking time more dynamically, roughly twice as fast on the fastest tasks and twice as slow on the slowest, and that the name reflects an iterative upgrade within the GPT-5 generation
Oct 15, 2025Anthropic — Claude Haiku 4.5 released — a small model at $1/$5 per 1M tokens, 73.3% on SWE-bench Verified (company-reported) SourceAnthropic said it gives coding performance similar to Sonnet 4, state of the art five months earlier, at one-third the cost and more than twice the speed; released under ASL-2 after testing found only limited CBRN risk
Sep 29, 2025Anthropic — Claude Sonnet 4.5 released — OSWorld 61.4% and SWE-bench Verified 77.2% (company-reported), with pricing unchanged at $3/$15 per 1M tokens SourceAnthropic said it observed the model maintaining focus for more than 30 hours on complex multi-step tasks; released under ASL-3 protections together with the Claude Agent SDK
Aug 5, 2025Anthropic — Claude Opus 4.1 released — an upgrade to Opus 4 on agentic tasks, real-world coding and reasoning SourceCompany-reported 74.5% on SWE-bench Verified; priced the same as Opus 4, with Anthropic recommending the upgrade for all uses
May 22, 2025Anthropic — Claude Opus 4 and Sonnet 4 released — hybrid models with near-instant and extended-thinking modes; Claude Code made generally available the same day SourceCompany-reported: SWE-bench Verified 72.5% for Opus 4 and 72.7% for Sonnet 4, Terminal-bench 43.2% for Opus 4. Priced like the previous generation ($15/$75 and $3/$15 per 1M tokens), with measures for higher AI Safety Levels such as ASL-3
Mar 6, 2025Alibaba Qwen — Released QwQ-32B — a 32B reasoning model trained by scaling reinforcement learning, under Apache 2.0 SourceThe full release following QwQ-32B-Preview; Alibaba claims performance comparable to DeepSeek-R1 (671B total, 37B active) at 32B
Feb 27, 2025OpenAI — GPT-4.5 released as a research preview — OpenAI’s largest chat model, built by scaling unsupervised learning, available to Pro users and developers SourceA non-reasoning model that does not think before responding; company-reported GPQA 71.4% and SWE-bench Verified 38.0% trailed o3-mini (high) at 79.7% and 61.0%, while it led on MMMLU (85.1% vs 81.1%) and SWE-Lancer Diamond (32.6% vs 10.8%)
Jan 31, 2025OpenAI — OpenAI o3-mini released — a cost-efficient small model in the reasoning series, with low, medium and high reasoning effort options SourcePreviewed in December 2024, it launched the same day in ChatGPT Plus, Team and Pro and in the API for usage tiers 3–5, and free users could try it through the 'Reason' button — the first reasoning model offered to free ChatGPT users. OpenAI cited particular strength in science, math and coding
Dec 11, 2024Google DeepMind — Gemini 2.0 introduced — first model, an experimental Gemini 2.0 Flash, with native image and audio output and native tool use SourceGoogle claimed 2.0 Flash outperforms 1.5 Pro on key benchmarks at twice the speed, launched the Multimodal Live API and Deep Research, and presented agent prototypes such as Project Astra, Project Mariner and Jules
Jul 18, 2024OpenAI — GPT-4o mini released — 82% on MMLU (company-reported), $0.15 input and $0.60 output per 1M tokens, 128K context SourceOpenAI said it is more than 60% cheaper than GPT-3.5 Turbo; it replaced GPT-3.5 for ChatGPT Free, Plus and Team users
Jun 20, 2024Anthropic — Claude 3.5 Sonnet released — the first Claude 3.5 model, $3/$15 per 1M tokens with a 200K context window, launched alongside Artifacts SourceAnthropic said it outperforms Claude 3 Opus at twice its speed and at mid-tier Sonnet pricing; on an internal agentic coding evaluation it solved 64% of problems (Claude 3 Opus: 38%). The UK AISI ran a pre-deployment safety evaluation, and Anthropic's red teaming concluded it remains at ASL-2
Mar 4, 2024Anthropic — Claude 3 model family (Haiku, Sonnet, Opus) announced — three tiers, vision input, 200K context SourceOpus and Sonnet available on launch day in claude.ai and the Claude API, generally available in 159 countries, with Haiku to follow. Per 1M tokens: Opus $15/$75, Sonnet $3/$15, Haiku $0.25/$1.25; assessed at ASL-2 under the RSP
①4 records
Prompt Engineering
Aug 9, 2023Anthropic — Claude Instant 1.2 released via the API — its faster, lower-priced model, now incorporating strengths of Claude 2 SourceCompany-reported scores: 58.7% on Codex (Instant 1.1: 52.8%) and 86.7% on GSM8K (80.9%); Anthropic said it hallucinates less and is more resistant to jailbreaks
Jul 11, 2023Anthropic — Claude 2 released — up to 100K input tokens per prompt, plus a public beta website, claude.ai, alongside the API (US and UK) SourceCompany-reported scores: Codex HumanEval 71.2% (Claude 1.3: 56.0%), GSM8k 88.0% (85.2%), and 76.5% on the multiple-choice section of the Bar exam (73.0%); API priced the same as Claude 1.3
May 10, 2023Google — PaLM 2 introduced — more heavily trained on multilingual text spanning over 100 languages, in four sizes: Gecko, Otter, Bison and Unicorn SourceAt Google I/O, Google announced over 25 products and features powered by PaLM 2, including Bard, and said the smallest size, Gecko, can run on mobile devices even offline
Feb 6, 2023Google — Bard announced — an experimental conversational AI service powered by LaMDA, opened first to trusted testers SourceGoogle said it would start with a lightweight version of LaMDA to reduce compute needs and gather feedback, before wider public availability in the coming weeks
③4 records
Tools & Agents
Sep 29, 2026OpenClaw Foundation, Red Hat, NVIDIA — OpenClaw Enterprise (OCE) released — an open-source, vendor-neutral platform for running always-on agents in sensitive environments. It provides multi-tenancy, hardened security boundaries, standardized agent primitives, and governance and auditability across the agent lifecycle, and is designed so that core components such as the harness, model and sandbox can be swapped for third-party or in-house ones. It began at OpenAI and was donated to the OpenClaw Foundation; it is co-developed with Red Hat and NVIDIA and being piloted at Red Hat and OpenAI. Says it runs on one's own infrastructure and is always free for any organization (no specific license name given). It can be self-hosted from the GitHub repository with docker-compose or Kubernetes, and 1.0 is planned for 2026. SourceA path has opened for agent runtimes to move from vendor products to foundation-owned standard infrastructure. The key is defining the harness, model and sandbox as swappable slots: if this separation takes hold, the cost of switching models falls and companies are less tied to any one lab. It meshes directly with demands from manufacturing and finance to keep always-on agents in isolated in-house environments.
Sep 29, 2026OpenAI — dots released (DevDay 2026) — an always-on agent built on GPT-6 Astra. It runs on its own cloud computer and connects to more than 4,000 apps through the OpenAI plugin ecosystem; users give it instructions and feedback via ChatGPT, Slack and Teams. The company says it works on its own from instructions and feedback and learns preferences over time. The Pro and Business Premium plans include a first dot at no extra cost, with expanded limits in the first month after launch. Enterprise can use it in beta once a workspace admin turns it on, and specialist dots that take on organization-level responsibilities were presented as well. SourceThe first major commercial product to sell an agent not as "a session that runs when asked" but as "a persona resident in the account". The billing unit shifts from tokens and seats to the number of agents, and always-on operation forces permission, audit and isolation design on the product side — touching, at a different layer, the same problem as the same week's OpenClaw Enterprise and Foundry egress controls.
Sep 28, 2026SpaceXAI (xAI) — Team Bots released — shared AI teammates built on Grok. A single bot bundles four things — context (files such as brand guides and documents, plus instructions and skills), plugins (Salesforce, Notion, GitHub, etc.), credentials (third-party API access) and memory (retaining what it learns so it improves over time) — and is shared by the whole team, while individual conversations stay private. It also collaborates in Slack through a dedicated bot handle, and comes with prebuilt templates for sales, product management, marketing and data analysis. Available the same day in public beta on Teams and Enterprise plans. The company is named "SpaceXAI LLC" in the announcement. SourceA design that treats agents not as personal assistants but as shared team assets. Bundling context, credentials and memory per bot means organizational knowledge builds up in the bot, but the bot in turn becomes a concentration point of permissions — the same direction as OpenAI's dots the same week, with greater security exposure because the credentials are shared.
Sep 24, 2026Microsoft — Routines in Microsoft Foundry Agent Service reach general availability (GA) — a managed way to run agents on a schedule, at a specified time, or in response to events in external systems. Without separately building schedulers, event listeners, webhooks or infrastructure, it gathers triggers, agent actions, identity, connections and run history inside the Foundry project. There are three kinds of trigger: timer (once, at a single future point), recurring (on a set cycle) and event-based (when something happens in a connected system). For identity, users choose between the creator's delegated permissions and the agent's own Entra ID. With a reminder tool, in preview, the agent can schedule its own re-run after a set time, monitoring multi-step tasks without an external polling service. The initial integrations are GitHub issue events and Teams channel messages. SourceThe operational foundation for agents moving from "running when a person calls them" to "waking up on their own" has become a GA cloud product. It matters that the option of giving agents their own identity came out at the same time: when the subject in audit logs is an agent rather than a person, the unit of accountability changes.
⑩3 records
Safety, Alignment & Interpretability
Sep 30, 2026OpenAI — Disclosed that it disrupted a coordinated model distillation campaign — activity from Jul 1 to Jul 28, 2026. Explains that it did not break into encryption or databases but was adversarial distillation that manipulated model interactions to reproduce protected reasoning in a form visible to the requester. It began at low frequency on July 1, and on July 24–25 more than 4,000 users made 16,000 requests with related extraction patterns; related activity spanned more than 15,000 users. Says it attributes the core cluster of activity to individuals linked to Moonshot AI, the developer of Kimi, but notes that attribution of the full set of operators is uncertain. Fully shut down on July 28 through account bans, restrictions on fraudulent accounts, tighter sign-up controls, expanded monitoring, blocking of reasoning-replay paths, added output-detection checks and coordination with third-party service providers. SourceThe first publicly disclosed large-scale extraction attempt since hiding reasoning traces became a trade-secret protection measure. Because the attribution points to a specific Chinese lab, it reads as a geopolitical event rather than a technical one, and the policy of keeping reasoning private gains a defensive rationale alongside the safety one. The attribution is OpenAI's claim, with no third-party checking.
Sep 29, 2026Anthropic — Published "GLM-5.3 and the spread of advanced cyber capabilities" — reports that GLM-5.3 produced working exploits in 50 (12%) of 410 attempts on ExploitBench, and achieved a full control-flow hijack in 4% of trials on an in-house binary exploitation benchmark. Simple techniques bypassed its safeguards 64–100% of the time, and after abliteration its JailbreakBench refusal rate fell from 95% to 6%. Concludes that GLM-5.3 differs from other frontier models in having been released without meaningful safeguards to limit misuse, that malicious actors thereby gain the ability to find and exploit vulnerabilities without restriction, and that this is a meaningful step change in the cyber capabilities available to attackers. SourceThe first case of a competing lab putting numbers on what leaks out when open weights are released without safety measures. The debate in the open-weight lineage (axis ⑨) shifts from a "performance gap" to a "safeguard gap", and the issue becomes not the release of weights itself but what was not released along with them. The interests at play when a competitor evaluates a competitor should be taken into account.
Sep 28, 2026OpenAI — Published "Towards safety cases for frontier AI training" — proposes producing a structured, evidence-based risk document (a safety case) before continuing frontier reinforcement-learning training. It writes that a safety case must cover all three aspects of the technical stack — alignment training, containment and monitoring — and that a senior accountable person responsible for the training run should be accountable for the safety case and incident response. It lists as components alignment through environment grading, infrastructure containment and sandbox hardening, real-time monitoring including automated responses, pre-mortems, an executive veto and audits, and publication of investigation findings. It describes these as best practices currently applied at OpenAI but, by its own account, "aspirational". SourceA record of the focus of safety discussion moving from deployment time to training time. If labs accept in public documents the premise that models can escape containment during training, a procedural standard emerges that regulation can refer to — and, taken together with Anthropic's GLM-5.3 analysis the same week, it also becomes a baseline against which to compare "releases without safeguards".
⑪3 records
Model Architecture & Training Methods
Apr 4, 2022Google Research — PaLM (Pathways Language Model) announced — a 540B dense decoder-only Transformer trained on 6,144 TPU v4 chips with Pathways SourceThe first large-scale use of the Pathways system; reported surpassing the few-shot performance of prior large models such as GPT-3, Gopher and Chinchilla on 28 of 29 English NLP tasks
Oct 11, 2018Google (Devlin et al.) — BERT paper posted on arXiv — pre-training a deep bidirectional Transformer encoder that conditions on both left and right context SourceFine-tuned with just one additional output layer, it set new state-of-the-art results on 11 NLP tasks, including GLUE 80.4% and SQuAD v1.1 F1 93.2 (as reported by the authors)
Jun 11, 2018OpenAI (Radford et al.) — The GPT (commonly called GPT-1) paper Improving Language Understanding by Generative Pre-Training — a 12-layer decoder-only Transformer pretrained on BooksCorpus, then fine-tuned for each task SourceReported that combining unsupervised generative pre-training with supervised fine-tuning improved the state of the art on 9 of the 12 tasks studied
⑧2 records
Inference Infrastructure & Serving
Sep 29, 2026Martin Hickey (IBM Research), vLLM — Published "Taking vLLM Apart: A Practical Guide to Disaggregated Serving" — sets out, with practical figures, prefill/decode disaggregation and moving CPU work off the GPU. On 2 L40S 48GB GPUs with Qwen2.5-7B-Instruct and prompts of about 8k tokens, the p99 inter-token latency (ITL) of a collocated deployment jumps from 23ms to 169ms at 0.4 req/s and reaches 263ms at 2 req/s, whereas prefill/decode disaggregation holds 25–52ms across all load levels. On AMD MI300X with Qwen3-235B, requests meeting both a 1-second TTFT and a 50ms ITL numbered 73/100 disaggregated vs 30/100 collocated. On multi-node H200, mean end-to-end latency fell 59% and P95 latency 67% compared with collocated replicas. Presents three kinds of separable work: (1) prefill/decode disaggregation (transferring about 3GB of KV cache over RDMA for a 10k-token Llama-3.1-70B prompt); (2) a GPU-less render layer that handles tokenization, chat templates and parsing (about 15ms of CPU for a 9k-token prompt); and (3) bidirectional KV transfer that reuses the cache in multi-turn conversations. SourceDisaggregated serving has come down from a research claim to an operational guide. The key is using p99 latency rather than average throughput as the yardstick, since agent loops are directly exposed to tail latency. It is the open-source side's response to the same pressure that led OpenAI, the same week, to split speed off into a separately billed tier.
Sep 29, 2026OpenAI — DevDay 2026 announcements — unveiled the Ultrafast speed tier: token generation up to 8 times faster in Codex (300 tokens per second) and up to 6 times faster in the API, available the same day in the API and on Pro 500 and Enterprise in ChatGPT Work/Codex. Also announced Private Intelligence (Zero Data Retention with Private Safety Processing — automated safety review runs without OpenAI staff seeing the content; a Private Inference preview in fall 2026), the Decisions API (real-time decisions on user-defined questions from a finite set of predefined answers), the Agents API including Computer Use, Bedrock Managed Agents, Codex Security Cloud (automatic checks of GitHub repositories and preparation of fixes), a revamped Codex CLI, ChatGPT Space, Pages and Team Tasks, a Pro 500 tier (25 times the Plus limits), and "Sign in with ChatGPT" with more than 16 partners. SourceThe key is that speed has been split off into a separately billed tier. When latency becomes the bottleneck of agent loops, tokens per second become a product separate from quality, and investment in inference infrastructure is recouped not through top scores but through latency guarantees by tier.
㉒2 records
Agent Security & Sandboxing
Oct 2, 2026Tencent AIG, OpenClaw — Tencent AIG joins ClawScan — announced that it will reinforce the security assessment of skills and plugins submitted to ClawHub with benchmark evaluation. SourceAs skill and plugin marketplaces grow, the supply chain becomes an attack surface. A structure that runs vetting through benchmarks from multiple participants rather than a single vendor follows the path package repositories have taken, and signals that MCP and Skills standardization (axis ④) is widening from distribution specs to vetting specs.
Sep 24, 2026Microsoft — Released network egress controls for Microsoft Foundry hosted agents — the destinations an agent may connect out to are defined on ARM resources as FQDN rules. Rules can specify four actions — Allow, Deny, request-header modification (Transform) and destination rewriting (Rewrite) — and policies run either in Audit mode, which only observes without blocking, or in Enforced mode, which actually blocks, with admins deciding whether unapproved destinations are denied by default. It targets situations in which an agent sends unexpected external requests, such as following an unfamiliar upload link in a document, or a new helper library that follows unexpected URLs; the design calls for a security boundary that can be reviewed separately from application authentication and authorization, rather than relying only on tool vetting. SourceA commercial implementation that moves prompt-injection defense down from model alignment to the network boundary. Starting from the premise that agents cannot be kept from being fooled, it keeps data from getting out even when they are, reaching at the product layer the same conclusion as the same week's NVIDIA and AgentKernel line of discussion. One more piece of cumulative evidence for the observation topic "agent security and sandboxing".
㉓2 records
On-Premises & Sovereign AI
Aug 18, 2026Ministry of Science and ICT — Announced the second-stage evaluation results of the independent AI foundation model project — Upstage, SK Telecom and LG AI Research advance; Motif Technologies eliminated. Scores combined benchmarks (40), expert review (35) and user evaluation (25) SourceAll four teams' models were judged to meet the independence criteria; with a newly introduced public user evaluation included, the field narrowed to three
Feb 20, 2026Ministry of Science and ICT — Selected an additional elite team for the independent AI foundation model project — the Motif Technologies consortium (with Moreh, Crowdworks, NdotLight, KAIST and Mathpresso); Trillion Labs, the other applicant, narrowly missed out SourceBrought the project back to four teams after the first-round cut, all to be evaluated in August under the same development period, GPUs and data
④1 record
MCP & Skills Standardization
Sep 29, 2026Jaewon Chu et al. — RASO (Retrieval-Augmented Skill Optimization via Cross-Harness Adaptation) released — a framework that optimizes agent skills by using existing skill repositories as base knowledge, rather than relying only on expensive agent rollouts. It runs in two stages: RASI builds initial skills grounded in existing knowledge without rollouts, and RASU retrieves external knowledge, guided by execution results, to improve skills step by step. Says it consistently beat baselines without retrieval-augmented initialization and updates across 4 agent benchmarks and 2 models. 16 pages. arXiv 2609.38024. SourceAn approach that retrieves and adapts skills from repositories instead of learning them anew for each harness, turning skills from outputs tied to a model and harness into transferable assets. Read together with ClawScan's tightened skill vetting the same week, the premise that skills are goods in circulation is firming up from both sides.
⑤1 record
Loop & Harness Engineering
Oct 1, 2026Sungho Park et al. (Microsoft and others) — ActiveSaddler released — points out that existing LLM-agent optimization deals with "how to update the harness" but largely leaves "which training scenarios produce that feedback" fixed, and adapts the curriculum together with harness evolution using a non-stationary bandit. It abstracts recurring agent failures into reusable failure patterns and estimates the learning potential of targeting each pattern, weighing exploitation of known weaknesses against exploration of new failure situations. Compared with baseline harness optimization using a fixed scenario order, +4.4 percentage points Pass@1 on GAIA2 and +7.5 points on Terminal-Bench 2.0. arXiv 2610.00906. SourceThe optimization target of harness engineering has widened from "what to fix" to "what to test". Abstracting failures into patterns to compose the curriculum automatically takes over the place where people used to pick scenarios in the agent-improvement loop, and touches on recursive self-improvement (axis ⑫).
⑫1 record
Training Data & Post-Training
May 18, 2021Google — LaMDA (Language Model for Dialogue Applications) unveiled — a Transformer-based conversation model trained on dialogue SourceUnlike most language models it was trained on dialogue, and Google said fine-tuning after training could significantly improve the sensibleness and specificity of its responses
⑬1 record
Compute Economics & Geopolitics
Oct 2, 2026Anthropic — Announced Claude Frontier Academy — an enterprise AI workforce program that will spend $100 million to train 10,000 Frontier Deployed Engineers (FDEs) by the end of 2027. Modeled on medical residency, participants first go through a multi-day in-person intensive course working through simulated enterprise deployments alongside Anthropic engineers, and those who pass a graded practical exam receive a Claude Resident Engineer badge. Next, in a 12-week residency, they lead a real Claude project at their own organization with support from Anthropic engineers, and on completion receive the Claude Frontier Deployed Engineer credential. The first cohort includes engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk, and runs in San Francisco, New York and London. The first graduates are expected in early 2027. SourceModel suppliers have begun issuing workforce credentials themselves. The diagnosis is that the bottleneck is not model performance but "people who can deploy it in companies", and if labs hold the credentials, the deployment staff of consulting and financial firms become tied to a particular model ecosystem. The point at which competition in compute economics spread from chips and power to skilled labor.
⑭1 record
Physical World & Real-World Impact
Oct 1, 2026Barclays, Anthropic — Disclosed Barclays's expanded use of Claude — the Colleague Knowledge Assistant, which UK staff use to find information relating to more than 20 million retail customers, is used by more than 16,000 people and has handled more than 1 million searches since its launch in 2025. The Global Markets organization uses Claude to triage and route about 120,000 incoming emails a day. Said it plans to extend Claude Code to 50% of its developer headcount by the end of 2026 and to the majority of its software engineers in 2027. Group co-COO Craig Bright said the bank is moving AI toward increasingly agentic capabilities embedded in the way it builds. SourceA rare case of a large bank in a regulated industry disclosing seat counts and processing volumes. 16,000 people and 120,000 items a day are operating scale, not a "pilot", and the target of 50% of developers shows coding-agent adoption turning from an option into a plan. All figures are the two companies' announcements.
⑰1 record
Embeddings & Vector Search
Sep 30, 2026Cohere — Embed 5 released — a family of embedding models in two tiers, Pro and Fast. Both tiers have a 128K-token context, variable output dimensions of 2048, 1536, 1024, 768, 512 and 256 (Matryoshka), more than 100 languages, text, image and fused text+image input, and float, int8 and binary embedding formats. Company-reported figures: ViDoRe V3 (enterprise documents) Pro 85.8 and Fast 84.5, FinanceBench 80.1, FinQA 90.0, ViDoRe Finance 85.0, fused text-image retrieval average 82.3, page-image financial dataset 77.0, and an average of 77 across German, French, Spanish, Italian and Russian (Farsi +13, Telugu +12, Hindi +12). The two tiers share an embedding space, so documents indexed with Pro can be queried with Fast without reindexing ("one index, two models"). Per 1M tokens: Pro $0.12, Fast $0.08 (images $0.40 for both). Available on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and North. SourceA case of the competitive focus of embedding models moving from scores to operating costs. The design in which tiers share an embedding space aims to eliminate reindexing costs — the biggest barrier to switching models in large-scale RAG — and if it works, the replacement cycle for embedding models will shorten.
㉔1 record
Diffusion Language Models
Oct 1, 2026Hui Ren et al. (UIUC) — HC-DLM (Hierarchical Continuous Diffusion Language Models) released — targets both the problem that discrete diffusion language models sample each token independently from its marginal distribution, cutting statistical dependencies, and the problem that continuous diffusion has no grounding in valid token configurations until final decoding. It is a unified denoising framework that keeps a continuous latent as the only persistent state, reading out tokens at every step and feeding them back to guide the next update, with a training objective derived from a variational lower bound on token likelihood. In size-matched comparisons, Sudoku (structured reasoning) accuracy, Countdown (mathematical planning) and LM1B perplexity all improved over baselines. arXiv 2610.02193. SourceThe point at which generation paths outside autoregression matured from the claim that "parallel generation is fast" to the question of "how to preserve dependency structure". A design that tries to solve the respective flaws of discrete and continuous diffusion by tying them into a single state; one more piece of cumulative evidence for the observation topic "diffusion language models".
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.