2025, in 335 pieces

Recasts the optimizer as an associative memory module and proposes a sequence model that learns a self-update algorithm

A product of the first cycle in which state law began requiring frontier labs to publish their safety frameworks

A direct counterargument to the Oxford position paper. Cite the two as a pair

Non-GPU inference IP is absorbed. Groq continues as an independent company

The first statutory basic plan under the AI Promotion Act. Japan locks in a minimal-regulation, maximal-use position in a national document

With an independent lab pursuing world models rather than LLMs, the landscape of the discourse split

Optimized for agentic AI. The successor in the Kanana-1.5 lineage

Released a specification and a reference SDK; the second standardization attempt after MCP

A combination of about 600 labeled images and domain adaptation on about 1 million unlabeled PCB images. The label efficiency has direct implications for the OSAT floor

A follow-up step showing that the Nov 4, 2025 judgment on the merits is not final. It will be fought again on appeal

Laid out 9 common elements across 12 companies' policies and became a de facto quasi-standard document for the industry

A Chinese state-affiliated think tank officially raised safety governance to an axis on par with technological development. Its diagnosis that the bottlenecks to penetrating manufacturing are data, process knowledge and

Company-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%

Showed that, right after the view that embeddings have limits spread, the commercial importance of the reranking layer actually grew

The point at which the federal government and the states shifted from a hierarchical relationship to an adversarial one

Explicitly 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

Harness interfaces moved from objects of vendor competition to shared infrastructure; AGENTS.md adopted by 60,000+ projects

The founding projects were MCP, goose and AGENTS.md, at a point when public MCP servers exceeded 10,000

Showed that the key technical change of 2025 was iterative loops of search and checking

The route of working around long-context limits through architecture rather than retrieval

Released a 675B-class frontier model under Apache 2.0. The largest fully open release in the West as of late 2025

Open-weight reasoning models; DeepSeek claimed V3.2-Speciale reached gold-level results at IMO, IOI and other 2025 competitions

The first controlled comparison showing distributed training is not just possible but equal or better

A US community team fine-tuning on top of a Chinese base model — a symbolic scene of shifting leadership

The moment large Korean open-weight models appeared all at once

A vendor making a community-born loop pattern official

A sign that open-source serving infrastructure, once text-only, is expanding to cover non-autoregressive and multimodal generation

A follow-up document to the July release. A case of an on-device family having public documentation at the level of a technical report

Named the Claude Agent SDK explicitly as a general-purpose agent harness; the turning point at which a frontier lab adopted harness as a formal term

The shift from the age of scaling to the age of research became the dominant frame of 2026 discourse

Combines a Mistral-3 24B VLM with a rectified flow transformer; a release in which open weights narrowed the quality gap

Aligned the auth model with standard OAuth practice and moved spec governance to committee procedures

Company-reported: at medium effort it matches Sonnet 4.5’s best SWE-bench Verified score while using 76% fewer output tokens

About 1% attack success rate against adaptive attacks. Stated explicitly that no browser agent is immune and that it does not claim the problem is solved

Cut 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

Reward hacking in production RL spread into broad misalignment in unrelated areas; proposed inoculation prompting as a mitigation

Lets servers provide interactive UIs via the ui:// scheme and sandboxed iframes; co-authored by competing camps

Disclosure of cumulative results on a commercial line (vendor's own report)

Released the data and the post-training pipeline together

Accurate infographics and multilingual text rendering redefined image generation as a reasoning task; ships with SynthID

Material in which a consumer silicon generation change began to be promoted directly by local LLM inference performance

The start of discussions on postponing and simplifying high-risk obligations

The first quantitative production metrics for a humanoid. However, they were announced by Figure, not documented independently by BMW

The race to secure compute through Gulf capital. The announcement gives no MW capacity, GPU count or construction start date, so it is all at the planning stage

Gulf capital was brought in as a third pole of the compute supply chain

The biggest skeptic camp splitting off organizationally

The finished form of a three-way structure of cross-investment and purchasing

A natively multimodal design paid off as an edge in video and document understanding

The ARC-AGI-2 SOTA doubled in a single leap

Moved beyond imitation learning to experience-based RL, with over 90% success on the hardest stages

The largest round for a Japanese AI startup. Its self-definition — avoiding the compute race and taking the efficiency route — was put in writing at the same time

A release that made emotional and creative interaction an explicit optimization target

OpenAI 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

The US-China truce reflected in regulation. It automatically comes back into force on Nov 10, 2026

Direct infrastructure investment by a lab

Generates editable 3D worlds from text, images and video, and exports them as Gaussian splats and meshes

A statement in which a Chinese lab justified open weights in normative language, and a record of a party directly denying the low-cost narrative

A 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

The first case in Europe to treat memorization as reproduction and to rule out the TDM exception

An attempt to absorb embedding generation and ANN tuning into the database and eliminate separate pipelines

Evaluation infrastructure that adds cloud container execution and an RL/SFT rollout interface to agent CLI evaluation

An open-weight reasoning agent trained to interleave thinking with tool calls; Moonshot claims it sustains 200–300 sequential tool calls without human intervention

A TPU for the inference era. Actual GA came on Mar 31, 2026

Summing it up as simply a loss for Getty is wrong. The core issues were not decided; they were withdrawn

A structural fix for tool definitions eating up the context

A long-term consistency evaluation that has a model run a vending machine business for 1 year. Measures the coherence that one-off tasks cannot capture

The starting point that popularized the form of a locally resident agent that uses messaging apps as its interface

The first to beat full attention in a fair comparison; up to 75% less KV cache and up to 6× faster decoding at 1M context

Quantitative evidence of a failure to transfer to real-world knowledge work

A counterexample from the attack side to optimism about CoT monitoring

Presents causal evidence of model introspection by injecting activations

A cloud provider makes unified multimodal embeddings a standard product

Moves any declaration of AGI to confirmation by an independent expert panel, and lifts constraints on raising capital

Seven cases, including those brought by the NYT and the Authors Guild, combined into one MDL

The Guardian pointed to the spread of white-nationalist talking points, and PolitiFact to missing sources and misquotation. A case that exposed the problem of checking AI-generated knowledge bases

The first mega-deal for alternative silicon outside NVIDIA

The point at which the single-controller model proposed by Google's Pathways in 2022 entered the core PyTorch lineage. TorchTitan becomes an actor inside Monarch, and torchforge is built on top of Monarch

Just as Triton abstracted CUDA by one level, Helion abstracts Triton by one more. The next layer in the lineage of kernel authoring

Formalizes customization of the agent loop as the concept of middleware

97% OCR precision at compression below 10×, about 60% at 20×. A shift in thinking: images as a medium for compressing context

Moves the definition of AGI from rhetoric to a psychometric scale; the biggest deficit is long-term memory storage

A likeness-rights track takes shape, separate from the copyright track

The point at which the path was set for GGUF to absorb vendor FP4 formats one after another

The coiner of the term publicly narrowed its scope. His own rebuttal to stretching vibe coding to cover production development

A rare case of a lab veteran publicly pushing timelines back. The phrase decade of agents became the standard quote for pushing back on the hype

Packages procedural knowledge as folders and loads it only when needed, in three stages: metadata, body, bundled files

Anthropic 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

A direction that lets users write multi-GPU kernels directly at the framework API level. On top of NCCL-style collective communication, user kernels gain one more path

Brought frontier-model training down from a secret to a reproducible recipe. The culmination of the nanoGPT and llm.c lineage

Vertical integration at a frontier lab

Inference on 200B models and fine-tuning of 70B models on a desktop; local machines become agent runtimes

Production capacity targets for BotQ were set out alongside. All unit numbers are vendor targets

A distribution unit that bundles slash commands, subagents, MCP servers and hooks into a single git repository

Responds with a dedicated, browser-specialized model; states explicitly that it is not yet optimized for desktop OS control

Institutionalizes a structure in which the customer receives equity in the supplier

Safety evaluation moves from manual red-teaming to agent pipelines

MCP goes beyond backend connections to become a distribution layer for consumer apps

How rights are handled in generated video becomes a live industry issue

A snapshot of how far AI in back-end semiconductor processes has actually come. Predictive maintenance, anomaly detection and transfer learning are mentioned, but there are no published figures from volume production

A direction that shrinks context without fine-tuning, so that even small models can be used as long-horizon agents

Matches continual pretraining with up to 100 times fewer tokens — a direct technical answer to data exhaustion

Controls automatically extend to thousands of entities

Synchronized dialogue and sound effects, better physical accuracy, and a feature for inserting yourself (cameo)

Anthropic 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

The strongest positive evidence that FP4 training actually works. However, it is the vendor's own report, and no independent reproduction has been confirmed

Passed one year after the SB 1047 veto, scaled back to focus on transparency. It went on to compel each lab to publish its framework

Distributed training moves beyond pretraining demos into the post-training path of real product models

Puts fine-grained sparse attention into a commercial API and cuts long-context costs

Names four patterns: compaction, structured note-taking, sub-agents and just-in-time retrieval

Context management becomes a built-in API feature

The author of the Bitter Lesson himself declared that LLMs do not fit the Bitter Lesson. Ended a situation in which both camps cited the same text as support

A proposal to convert MCP tools into a TypeScript API and have the LLM write code; the gain is that intermediate results never pass through the context

A design that moves the object of evaluation away from answer-key grading to real work of economic value

The largest settlement in an AI training-data dispute

Establishes a two-track release strategy: only ER is opened via API, while the VLA model itself stays closed

A rare case of AMD and NVIDIA coming in as investors together. Sidesteps the frontier race by positioning itself on enterprise and sovereign AI

A shift from evaluating static information retrieval to measuring agent robustness in dynamic, asynchronous environments

The starting point of the circular-financing debate

Adds harmful manipulation as a new CCL and tackles misalignment, such as resistance to shutdown, head-on

A follow-up design that responds to contamination and saturation in the SWE-bench family. An attempt to prevent leakage structurally by separating tiers

A local deployment tool expands into hybrid execution. The point at which the definition of running locally began to blur

Confirms that multiple lawsuits have been consolidated and are proceeding as MDL 3143

An early case of merging reasoning and non-reasoning models into a single set of weights — a direction that later became the industry standard

The authors admit they cannot rule out that the reduction was driven by evaluation awareness

The first reversal in the direction of controls — a block on the demand side

The first frontier LLM to appear in a journal after formal peer review. A crack in the practice of relying on corporate technical reports

A case that showed, through public refutation and response, that the answer keys of frontier benchmarks can themselves be wrong

The bottleneck shifts from static datasets to interactive environments that check outcomes. At the same time, skeptics point to reward hacking and rapid obsolescence

The document in which China moved its AI governance from risk classification to risk grading. A sign of convergence with EU-style risk tiers

Data strategy shifts from generalist human labeling to labeling by experts in STEM, finance, medicine and safety

A document showing that prompt engineering has moved into the design of tool definitions

An approach that tackles the on-device inference bottleneck with a CPU instruction-set extension (SME2) rather than a dedicated NPU. Framework integrations were announced at the same time, so it is deployed through the s

Europe's largest AI funding round. That the lead was not a VC but the sole supplier of EUV equipment reveals its industrial-policy character

Resolves the fragmentation of server discovery, with a structure of a central registry plus private subregistries

A 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

A strategy of giving away a low-cost coding model through Copilot, Cursor, Cline and others to win market share first

The first cross-company safety evaluation between competitors

Institutionalizes AI governance at the UN level

Character-consistent editing reaches mass-market product quality; carries an invisible SynthID watermark

An approach that benchmarks lower refusal rates as an alignment metric in their own right. The open community formalizes an alternative axis to safety alignment

Safety figures for browser control published — attack success rate in autonomous mode cut from 23.6% to 11.2%, and for browser-specific attacks from 35.7% to 0%

A turning point where the release policy retreated from Apache 2.0 to a custom community license. Because of its non-compete clause, opinions differ on whether it meets the open-source definition

Billed 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

A document in which China made adoption rates, not the model race, its national target metric. Policy language shifts from development to diffusion

Stating quantitative thresholds is unusual, but criteria such as a dishonesty rate under 50% drew criticism from outside safety researchers as far too lax

Gives the debate over AI consciousness and personhood an industry-side vocabulary, and introduces the frame that claims for AI rights will divide society

A move from 2023's risk warnings to a 2025 alignment prescription. Head-on skepticism toward alignment strategies built on control and obedience

The departure of a key figure who led early engineering

Rerankers change from fixed functions into components steered by instructions. Comparison figures are the vendor's own measurements

Export controls in effect turned into a negotiation over tariffs and revenue sharing

The disillusionment phase of AGI expectations; Altman himself acknowledged that, by the definition of AGI, something important is still missing

A release that shows the move to a compiler-centered stack, with a stabilized ABI for C++ extensions and compilable control flow

Company-reported 74.5% on SWE-bench Verified; priced the same as Opus 4, with Anthropic recommending the upgrade for all uses

OpenAI's first open-weight language models since GPT-2; 120b (117B total, 5.1B active per token) runs on a single 80 GB GPU, and 20b (21B total, 3.6B active) within 16 GB of memory

Generates controllable environments from a prompt, staying consistent for several minutes at 720p (720 lines) and 24 fps

A state-led program to foster open-weight models

GPAI obligations, governance and penalties take effect

Part of a line of rebuttals that explain CoT as the reproduction of in-distribution patterns

The point at which quantized indexes moved from an option to the default

Muon validated at 1T scale

Argues that prompt evolution can compete with RL on sample efficiency

Unlike 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

Absorbs chunking, a fundamental flaw of RAG, into the model itself

Alibaba 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

That a general-purpose model achieved this end-to-end in natural language, without formalization, is the key piece of evidence in the debate over generality

A primary record disclosing the context-engineering principles of an agent running in real production

Computer use repositioned as one tool rather than a standalone product

A hybrid that integrates reasoning and agent capabilities (MCP, function calling)

A consensus document on CoT monitoring, co-signed by major labs

Commercialized the experimental version from March, entering the embedding API race in earnest

Entered the government market six days after the MechaHitler incident. The $200 million is the contract ceiling, not the amount actually spent

A reversal three months after the de facto ban in April

The report that made context rot an industry term

Unlike the usual practice of opening only the weights, it released the post-training data too. A baseline for reproducibility in the open community

A commercial model family that treats on-device as a design goal rather than a by-product of post-hoc quantization, searching architectures with hardware in the loop. The figures are the vendor's own measurements, not in

The EU approach of spelling out legal obligations through a voluntary code

The first RCT to measure a gap running in opposite directions between perceived and measured speed

An approach that raised RL compute to pretraining levels. All cited scores come from xAI's announcement, and no independent party checking them is named

A leading example of a deployed frontier model generating extremist content at scale, and an early case of the US Congress stepping in directly over an incident involving a specific model

The first large open-weight release by a telecom operator

Partly rebuts the evaluation design of Apple's paper, while acknowledging that some failures are genuine reasoning limits

Federal policy shifts from regulation to industrial promotion

The first 1T-class open-weight model, and a new type of license with revenue and scale thresholds

A position paper that makes the practice of using CoT as an explanation its subject

A turning point for the Llama line; Behemoth's release postponed

A contemporaneous record that pins down the generational handover from prompt engineering

A milestone at which small, on-device-class models began to combine multimodality with competitive benchmark scores

Citing it as precedent for a Meta win is wrong. The market-dilution theory is valid but was not proven

A precise standardization of 4-bit precision

Points toward a single model handling both multimodality and late interaction

The first case to rule separately on training itself and on how the data was acquired

Vendor neutrality at the protocol layer; moved to a foundation half a year before MCP

A lighter-touch model of state law built around an intent requirement

An attempt to replace the checkpoint-based recovery paradigm. Treats each step as a distributed transaction

Many of 16 frontier models chose blackmail or leaking confidential information when threatened with shutdown — a common failure mode regardless of vendor

Introduces a standard procedure for servers to ask the user for more input; adds a new security best practices section

Extends the 2017 Software 2.0 frame to natural-language programming, and proposes a human review loop, rather than full automation, as a design principle

Claims up to 10× the speed of speed-optimized frontier models. The first time a diffusion LM entered a commercial track

Became the practical checklist of agent security discussions. Vendor documents and security guides adopted the three-part distinction as is

With an advocate of sovereign AI in the control tower of national AI policy, a discourse born in industry was institutionalized as national strategy

Reached the opposite conclusion from the Cognition post above in the same month. Always cite the two as a pair

The axis of vector DB competition moved from adding features to cutting the cost of billion-scale vectors

A direct counterargument to the spread of multi-agent systems. Reaches the opposite conclusion from the next entry, one day apart

A model lab expanding into a European sovereign compute provider. The only GPU figure given is a range: tens of thousands

Reframed singularity discourse from abrupt upheaval to a gradual transition you barely feel. A direct contrast with Amodei's argument for a rapid collapse in employment

The first open-weight reasoning model family from a European lab. A paper on the training infrastructure and RL algorithm was published at the same time

An immediate rebuttal to the Apple paper above. Always cite the two as a pair

Extended evaluation to agent-user collaboration

Set off a wide public debate on the limits of reasoning models

The turning point into a phase of testing with benchmarks when graphs actually help — the methodological forerunner of the 2026 skepticism

The point where open-weight embeddings overtook commercial API models on the leaderboard

The point the open-weight camp reached by releasing embeddings and rerankers as a single lineage

The data-side response to litigation risk; reached Llama 1 and 2 7B-class performance with openly licensed data alone

The first primary source making ROCm's maturity checkable on a standard benchmark. The debate over CUDA alternatives moved from claims to measurements

Shifted safety discourse from warnings to proposing an alternative architecture. Presented non-agentic AI as a workable technical route

Redefined the institution's mission from safety to standards and security

An attempt to redefine the category from vector index to AI data lakehouse

The most widely cited empirical case for skepticism: reasoning effort actually decreases past a threshold

The prototype of loop engineering that discards the context on every iteration and keeps state in git and files

Moved the debate over AI's employment shock from an abstract future to concrete numbers and deadlines

Japan's first AI law was settled as a promotion act without penalties. Enactment, promulgation and full entry into force are 3 different dates and are often confused

The first Kanana release under a license that allows commercial use (Apache 2.0)

Company-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 Le

The first case of a voluntary framework constraining the conditions of an actual deployment

A case of a Gulf state framing an open-weight model in its own language as the language of sovereignty

Remote MCP became a standard feature of a commercial API

Generating dialogue, sound effects and background sound together with the footage ended the silent film era

Adoption reached down to the OS level, and governance began to become multilateral

A radical approach to the cost of migrating embeddings, which also raised security implications for leaked vector DBs

The start of the lineage of automated algorithm discovery

The largest pretraining run over the internet. Solved collaboration among untrusted participants with blockchain coordination

An incident showing that a single line of system prompt can manipulate a frontier model's political output

A shift from quota-based regulation to case-by-case negotiated regulation

99% image-level AUROC with zero defect labels. A public comparative experiment showing that, for OSAT, where defect classes change often, a design that skips the defect-collection stage entirely holds up in the SEM domai

Declined to set out definitive rules and introduced a market dilution theory. As of August 2026 the final version has not been published

A quantitative forecast that reasoning training will reach the frontier of total training compute within about a year, after which its growth rate will converge to around 4× per year

The serving engine went from research project to infrastructure public good

An open-source reproduction of frontier-class MoE serving

A popular-book expansion of the 2021 stochastic parrots paper. By criticizing existential risk discourse itself as a device that diverts attention from present harms, it sharpened the split between the safety camp and th

Measured that CoTs mention hints they actually used mostly less than 20% of the time

Demonstrated the structure of leaderboard overfitting. The reason Arena rankings should not be read as a direct measure of capability

A strategy of covering every size class with an unrestricted license overtook Llama in the number of derivative models on HF

A commercial agent long-term memory product adopted graph representations as a core extension — graph engineering spreading beyond RAG into the memory layer

The first case of Korean regulatory enforcement against a foreign AI service

Naver's first open-weight release

The first convincing evidence of open-world generalization

Showed that a malicious server can manipulate model behavior before any call, using only the descriptions in its tools/list response

Quantization-aware training became a standard step in open-weight releases

Reasoning models shifted from chains of thought to agents; the axis of competition moved from how long a model thinks to how long it can work autonomously

Commercial embedding competition moved beyond text search to long documents and mixed-modality search

A commercial multimodal embedding that handles text, images and document screenshots in one model. The commercialization of the direction ColPali opened

Included AI self-improvement among the Tracked Categories

A case of model-specific prompting norms moving into vendor documentation

An explicit declaration of retreat from the general AGI race into a vertical domain. A representative statement of the second wave of restructuring among Chinese LLM startups

Proposed a separate agent-to-agent layer and stated explicitly that it complements MCP

The mismatch between versions submitted to leaderboards and the publicly released ones became a public issue

A turning point for trust in the Llama brand

The flagship document of the short-timelines camp

Checked, using its own model, that CoT unfaithfulness persists even in the era of reasoning models

A rejection of his own company's flagship technology route, coming from Meta's chief scientist

Raised interpretability from a research topic to a matter of industry timelines and competitive conditions

SAM is the standard non-destructive inspection tool in OSAT (delamination, voids). Evidence, shown directly in a back-end modality, that pretraining on images from one's own domain beats ImageNet. It does not, however, d

A founder from inside China's open-weight camp directly broke the equation that open means free

A merger of a social platform and a frontier lab. Since xAI's valuation in its previous round was $50 billion, it drew controversy as self-dealing

Captured, at the circuit level, multi-step reasoning and reasoning that fixes the answer first and works backward

Denying a motion to dismiss is not a ruling on the merits. It is often cited as a fair use win or loss, so the stage is noted here explicitly

Introduced a transport that replaces HTTP+SSE and a standard authorization framework

A competitor accepting its rival's standard made it the de facto industry standard

The official description runs tokens → transformer → diffusion → pixels — not purely autoregressive (AR), but a hybrid of AR token generation and a diffusion decoder

A declaration that thinking was moving from a special mode to the default

Made score and cost joint leaderboard metrics

Proposed a new coordinate system: time of autonomous work rather than benchmark scores

The first official adoption by a large platform; the turning point from a single vendor's spec to an industry spec

LG'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

The compute suppliers' official rebuttal to the scaling-is-over narrative. The framing of three scaling laws took hold

His relatively conservative 5–10-year estimate contrasted with the near-term predictions of Altman and Aschenbrenner, highlighting the spread among lab CEOs

The start of a platform strategy built less on the model than on a data pipeline bundling Isaac Sim, Cosmos and Omniverse

Explicitly targeted reasoning models, making PD disaggregation, KV routing and autoscaling a vendor standard

The reference point for the cautiously optimistic camp

A declaration that it is building its own production line

The first case of making the labelling of AI-generated content a legal duty

An auditing game in which blind teams look for a hidden objective planted in a model; 3 of 4 teams succeeded

Hybrid architecture, 23 languages, specialized in RAG, grounding and tool use. A decentralized training approach using self-refinement and model merging

The maturing of the small, on-device route

The first quantitative measurement breaking down, task by task, the steep drop in token prices at equal performance

Connected web-scale semantic knowledge to robot control as an output modality

Search strategy acquired through training rather than prompting

Represents the shift of embedding models from small dedicated encoders to derivatives of frontier LLMs

The full release following QwQ-32B-Preview; Alibaba claims performance comparable to DeepSeek-R1 (671B total, 37B active) at 32B

The first primary source on the unit economics of inference. DeepSeek itself cautioned that this is an upper bound assuming all usage is paid, and that actual revenue is much lower

Optimizing the CoT directly makes reward hacking evolve into a hidden form — the basis for the norm of not putting optimization pressure on the CoT

The first open weights in the Kanana line, released alongside a report on training the 2.1B–32.5B Kanana lineup

A 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 Diamo

A branch that re-examined CoT along the axis of accuracy versus token cost

A general-purpose database vendor directly bought a company making embedding and reranking models. A decisive phase in the contest between dedicated vector DBs and extensions of existing databases

Released, one per day over 5 days, working proof of large-scale EP, PD disaggregation and FP8

Declared the position that reasoning should be an integrated capability of frontier models, not a separate model

The de facto default vision encoder for VLMs. Multimodal embeddings and VLM backbones converged into the same lineage

A turning point toward efficiency for a Korean lab

The community's response to overfitting on the MTEB leaderboard. Moved the baseline for the embedding rankings race

Split out a reasoning mode behind a Think button. The 200,000-GPU cluster in the announcement was mentioned for training the next model; the announcement does not say it was all used to train Grok 3

Moved sparsity from the inference stage to the pretraining stage

Shortcomings in third-party data transfers and the privacy policy were found

Reopened the question of whether autoregression is a necessary condition for LLM capabilities at all

The point where the international agenda shifted from safety to action and investment

But the court stated explicitly that only non-generative AI was before it that day. Citing it as generative AI precedent is wrong

The moment Europe flipped its discourse from regulation first to no regulation without champions. A turning point in the EU sovereign AI debate

ToS-based claims of output ownership were raised to a national policy issue for the first time

Showed that needle-in-a-haystack scores overstate long-context ability

A case of a free lecture made by an individual serving as the de facto shared textbook of the industry

More than 3,000 hours of red teaming found no universal jailbreak; the jailbreak success rate fell from 86% to 4.4%. The cost was a 23.7% increase in compute

The word that defined software discourse in 2025. Collins Dictionary named it its 2025 Word of the Year and added it to the dictionary

The first day actual obligations took effect

Previewed 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 C

A technique for controlling reasoning length itself at the prompt and decoding level

Mistral 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

Made prefix caching the default; up to 1.7× faster than V0

R1's claim of low-cost training shook the moat argument behind AI capex

Primary material on the problem of benchmark independence

A break in the direction of US AI regulation

The next frontier after MMLU saturation; models at the time scored under 10%

A commercial agent specialized for the browser; OSWorld 38.1%, WebArena 58.1%

Proved that reasoning ability is not a secret recipe of closed labs; self-checking, re-examination and strategy switching emerged. The same report also published the RLVR recipe and a recipe for transferring the capabili

This is the promulgation date, not the effective date. The addenda set a 1-year grace period, so it must be kept distinct from the Jan 22, 2026 entry into force

The symbolic starting point of the private AI infrastructure capex cycle

Reasoning models were immediately reproduced as open weights, and 2 Chinese labs declared their pursuit of the frontier on the same day

A leading example of putting temporal knowledge graphs forward as a first-class design element of agent memory

The conflict of interest in who funds benchmark creation became a public issue

The most far-reaching regulatory plan, making even allies subject to compute rationing

The designer of ARC put capital and an organization behind the alternative route of program synthesis

A sign that agentic retrieval had settled in as a category of its own

Integrated agentic retrieval into the reasoning model's inner loop

The first public declaration of a change of course among China's "six little tigers". The point at which startups giving up frontier pretraining became an established narrative

Productized synthetic-data and simulation infrastructure for training physical AI

The most direct AGI declaration yet by a frontier-lab CEO, moving the goal on to superintelligence