AI Timeline.

source-checked milestone · 16 Jan 2013 · Mikolov et al. / arXiv

word2vec makes word vectors cheap to train

Mikolov, Chen, Corrado and Dean described two efficient methods for learning continuous distributed vector representations of words from large text corpora, and reported that the resulting vectors captured semantic and syntactic regularities evaluated by analogy questions.

Why it matters

Test whether a trained word-vector model already separates the vocabulary your classification or search task depends on before building a bespoke representation layer.

Limits

This is a research preprint, not a released product, hosted service or benchmark suite. The reported regularities are the authors' own demonstration on their corpora and were not reproduced here, and analogy accuracy is not a measure of fitness for any particular task. It did not invent distributed representations: the paper extends earlier neural language modelling, including the approach this timeline records as rejected for its 2003 date. No Australian availability question arises. Prepared with AI assistance.

Read the decision guides

• Choosing AI: companies, products and models

source-checked milestone · 10 June 2014 · Goodfellow et al. / arXiv

Generative adversarial networks pair two models against each other

Goodfellow and colleagues proposed training a generative model by pitting it against a discriminative model that learns to tell its samples from real data. The paper reports that the approach trains a generator through that competition rather than through an explicit likelihood objective, and demonstrates samples from several image datasets.

Why it matters

Before commissioning a bespoke generative model, check whether an adversarial or likelihood-based training objective fits the data and the evaluation you can actually run, and decide in advance how you will judge samples you cannot score automatically.

Limits

This is a research paper, not a product, hosted service or benchmark suite. Quantitative evaluation in the paper is limited to Parzen-window log-likelihood estimates on small datasets, which the authors themselves describe as having high variance and not being the whole story; sample quality is otherwise shown visually. Those estimates are the authors' and were not reproduced. Adversarial training is not the only generative approach and the paper does not claim it is. No Australian availability question arises. Prepared with AI assistance.

Read the decision guides

• Choosing AI: companies, products and models

source-checked milestone · 1 Sept 2014 · Bahdanau, Cho and Bengio

Attention enters neural machine translation

Bahdanau, Cho and Bengio described a neural machine translation model that learns to search the parts of a source sentence relevant to each word it generates, instead of compressing the whole input into one fixed-length vector. On English-to-French translation the authors report performance comparable to the phrase-based statistical system then treated as the state of the art.

Why it matters

Read a single fixed-length encoding of an input as a documented bottleneck with a known alternative: where translation, summarisation or extraction quality and traceability matter, check whether the architecture can expose per-token source alignment you can inspect.

Limits

This is a research paper, not a product, and it records no released software or service. Attention here is the soft-alignment mechanism the paper describes; the Transformer is a separate 2017 entry that followed, and the paper does not claim to be the first attention mechanism. The reported translation quality is the authors' comparison on one language pair against one phrase-based system, not an independent reproduction or a general benchmark. Australian access does not arise for a paper. Prepared with AI assistance.

source-checked milestone · 10 Dec 2015 · He et al. / arXiv

Residual networks make very deep image models trainable

He, Zhang, Ren and Sun described residual learning, in which each block learns a residual function over a shortcut connection, and reported that networks with more than a hundred layers could be trained and still improve accuracy. The paper reports first place in the ILSVRC 2015 classification task with an ensemble and a 3.57% top-5 error on the test set.

Why it matters

Treat depth as an engineering choice with a known failure mode: if a deeper model performs worse on your data, look for a representation that carries information around the block before concluding the task needs more parameters.

Limits

This is a research paper; competition results are the authors' reported numbers, and the winning entry used an ensemble, so the single-model figure should not be read as the headline result. Comparisons with earlier very deep plain networks are the paper's own experiments. No independent reproduction, no Digital Sanctum test and no Australian availability question arise. Prepared with AI assistance.

Read the decision guides

• Choosing AI: companies, products and models

source-checked milestone · 15 Mar 2016 · Google DeepMind

AlphaGo wins a five-game match against Lee Sedol

AlphaGo completed a five-game public match against Lee Sedol, winning four games to one. DeepMind described the result as a major milestone for artificial intelligence, and recorded that it would donate the prize money to UNICEF, STEM charities and Go organisations.

Why it matters

Read a headline match result as a demonstration of a particular system on a particular day, not as a general ranking of AI against people: check which rules, time controls and versions were used before treating it as evidence for any other task.

Limits

This records the public match against Lee Sedol, not the earlier 2015 match against Fan Hui, and not a product or service. Commentary about a decade-earlier-than-predicted milestone is DeepMind's characterisation, not an independent forecast. No Digital Sanctum reproduction of the games was performed, and no Australian availability question arises. Prepared with AI assistance.

source-checked milestone · 12 June 2017 · Vaswani et al.

Attention Is All You Need

The original paper introduces a different architecture for sequence transduction.

Why it matters

Evaluate attention-based sequence models as an alternative architecture for language-processing work.

Limits

This is a research foundation, not a hosted product or an autonomous-agent release. Later paper revisions are not new milestones.

Read the decision guides

• Choosing AI: companies, products and models
• Test AI against the work, not the flagship label

source-checked milestone · 11 June 2018 · OpenAI

GPT-1 pairs Transformer pre-training with fine-tuning

OpenAI published a two-stage recipe that pre-trained a Transformer on unlabelled text with a language-modelling objective, then fine-tuned the same model for individual tasks, reporting state-of-the-art results on most of the dataset results in the table it published. The announcement presents the work as a combination of two existing ideas, Transformers and unsupervised pre-training, and compares its scores with the approaches it builds on.

Why it matters

Test whether one general pre-trained model plus a small supervised set covers a task before commissioning a bespoke system, and check the tasks where the reported results did not lead rather than assuming uniform improvement.

Limits

Every score is OpenAI's own comparison table on the listed datasets, not an independent evaluation, and the same table shows the system behind the previous best on tasks such as RTE. The post names Semi-supervised Sequence Learning and ULMFiT as prior work and does not claim a new principle. It is a research release: no API, product or weight release is recorded here, and no Australian access, commercial availability or Digital Sanctum test is established. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Choosing AI: companies, products and models

source-checked milestone · 11 Oct 2018 · Devlin et al. / arXiv

BERT pre-trains a bidirectional Transformer encoder

Devlin, Chang, Lee and Toutanova described BERT, which pre-trains a Transformer encoder on unlabelled text using both left and right context and then fine-tunes the whole model for each task with an added output layer. The authors report state-of-the-art results on eleven language-understanding tasks, including pushing the GLUE benchmark to 80.4.

Why it matters

Weigh a pre-trained bidirectional encoder as a reusable component for classification, extraction and question answering before training a task-specific model from scratch, and check whether your task is among those actually measured.

Limits

This is a research paper submitted to arXiv, not a released product; later code and weight availability is a separate event that this row does not record. All scores are the authors' claims in the first version's abstract on the listed benchmarks and were not reproduced here; later revisions report slightly different figures, so the row is bound to the version dated 11 October 2018. Bidirectional pre-training builds on earlier representation learning, and the paper does not claim to have invented pre-training or the Transformer; its comparison is with the representation models that preceded it. No Australian availability question arises, and no Digital Sanctum test was run. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Choosing AI: companies, products and models

source-checked milestone · 14 Feb 2019 · OpenAI

OpenAI withholds GPT-2 and stages its release

OpenAI described GPT-2, a 1.5-billion-parameter language model trained on 40GB of internet text, and stated that it was not releasing the trained model because of concern about malicious applications. As an experiment in responsible disclosure it released a much smaller model for researchers, together with a technical paper.

Why it matters

Treat the decision to withhold a trained model as distinct from the decision to publish a paper: if you depend on weights, check which checkpoint is actually available and on what terms, rather than assuming a research announcement includes them.

Limits

The announcement states OpenAI's own reasoning; it is not independent evidence that withholding reduced harm. Later stages of the same page, including a May 2019 interim update that released a larger checkpoint and shared still-larger ones with partners, are separate events and are not recorded here. Benchmark and sample claims are the publisher's and were not reproduced. No Australian availability is established. Prepared with AI assistance.

Read the decision guides

• Hosted access versus operating weights: who does what?
• Read the licence attached to the exact weights
• Open weights, closed weights and the rights your firm needs

source-checked milestone · 22 May 2020 · Lewis et al. / arXiv

Retrieval-augmented generation pairs a model with a document store

Lewis and colleagues described retrieval-augmented generation, which pairs a pre-trained sequence model with a dense index of documents so that output is produced from retrieved passages rather than from model parameters alone. The authors report state-of-the-art results on open-domain question answering and describe the index as editable, so knowledge can be updated without retraining.

Why it matters

Separate what a system must remember from what it must reason about: where answers depend on material that changes, test a retrieval layer you can inspect and update instead of retraining, and check that the evidence it retrieves is evidence you would show a client.

Limits

This is a research paper, not a product or hosted service, and its benchmark results are the authors' claims on academic datasets rather than an operational evaluation. Retrieval does not guarantee accuracy: a stale or wrong document can produce or contribute to a wrong answer, and the paper does not measure that failure mode for business use. Retrieval over external knowledge and related architectures predate this work. No Australian availability question arises. Prepared with AI assistance.

source-checked milestone · 28 May 2020 · OpenAI researchers / arXiv

GPT-3 shows in-context learning at scale

Brown and colleagues described GPT-3, a 175-billion-parameter autoregressive language model, and reported that it could perform many tasks from an instruction and a few examples supplied in the prompt, without gradient updates or task-specific fine-tuning. The paper presents that few-shot behaviour as distinct from the model's zero-shot scores.

Why it matters

Test whether a task can be specified in the prompt before commissioning fine-tuning or a bespoke pipeline, and record the prompt, the examples and the model version, because the same model can behave differently under a different framing.

Limits

This is a research paper about a model that was not generally released at publication; the separate 11 June 2020 API entry records a different decision and this row does not update it. Capability figures are the authors' evaluations on their own task suite, not an independent benchmark, and they do not establish reliability for any business process. In-context adaptation builds on earlier prompting and few-shot work. No Australian availability question arises for the paper. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Choosing AI: companies, products and models

source-checked milestone · 11 June 2020 · OpenAI

General-purpose language models through an API

OpenAI opens applications for a general-purpose text-in, text-out API using GPT-3-family weights.

Why it matters

Consider applying for hosted model access to build a language feature without operating model infrastructure.

Limits

An application path is not immediate access. Australian eligibility and suitability for client data are unresolved.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 10 Nov 2020 · Apple

Apple's M1 puts the GPU and CPU on one memory pool

Apple announced the M1 system on a chip, which the company describes as having a unified memory architecture: high-bandwidth, low-latency memory in a single pool inside the package rather than a processor pool plus separate graphics memory. For local inference the practical consequence is that the GPU addresses the system pool, so what fits depends on the memory the runtime can use and on what remains after everything else the machine is doing, not on a separate graphics-memory allocation.

Why it matters

Check which memory the runtime can actually address: on this design the GPU shares the system pool, so model size is bounded by the usable share of total memory rather than a dedicated card allocation, while speed still depends on bandwidth and on the kernels the runtime ships.

Limits

Apple's announcement gives no inference benchmark, and none is claimed here. Unified memory is not new: integrated graphics, consoles and other systems on a chip already shared memory, and Apple's own fine print in the same release defines an integrated GPU as sitting behind a unified memory subsystem. Capacity is not bandwidth and not measured throughput. The claim is confined to this SoC's memory arrangement, not to the origin of the term or to model performance. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Local AI memory budget worksheet: estimate first, then measure

source-checked milestone · 30 Nov 2020 · DeepMind / CASP14

AlphaFold 2 is assessed as competitive in CASP14

The CASP14 organisers announced that DeepMind's AlphaFold 2 produced protein structure predictions competitive with experimental structures in the blind assessment, for roughly two-thirds of the assessed targets, which were single proteins or domains. DeepMind described the result as a solution to a fifty-year-old grand challenge in biology; the method was published later.

Why it matters

Treat a benchmark result as a claim about one task under one assessment: before adopting predicted structures in research or client work, check the confidence measures, the assessment's own caveats and whether the prediction has been experimentally confirmed.

Limits

The programme's claim of competitiveness is the assessors' summary, not an independent replication, and predictions are not experimental structures. Scores and rankings are reported by CASP and DeepMind and are not reproduced here. 'Grand challenge' is DeepMind's characterisation. The later peer-reviewed method paper and the subsequent database are separate events. No Digital Sanctum test was performed. Prepared with AI assistance.

source-checked milestone · 29 June 2021 · GitHub

Copilot brings suggestions into the coding workflow

GitHub launches a limited technical early access of its AI pair programmer.

Why it matters

Evaluate suggestions made while a developer writes code, with the developer reviewing each proposed change.

Limits

This is assisted coding, not permission to run a repository unattended. The February 2022 page update is not another launch.

Read the decision guides

• Longer-running coding agents: when to let them continue
• Web, CLI, IDE, API or messaging: choose an interface
• Before staff use an agent on a real repository
• Test AI against the work, not the flagship label

source-checked milestone · 22 July 2021 · DeepMind and EMBL-EBI

AlphaFold Protein Structure Database opens to researchers

DeepMind and EMBL-EBI released the AlphaFold Protein Structure Database, making more than 350,000 predicted protein structures, including the entire human proteome, freely available to the scientific community. The partners described it as the most complete database yet of predicted human protein structures.

Why it matters

Treat a predicted structure as a starting point for experiment design, not as experimental confirmation: check the model's own confidence measures and the licence terms before relying on a record in a paper, grant or client report.

Limits

Predictions are not experimental structures. The announcement's completeness and accuracy claims are the publishers', covering the predicted human proteome as of 22 July 2021, not every sequenced protein then known. Later expansions of the database are separate events. No Digital Sanctum download, structure check or Australian access test was performed. Prepared with AI assistance.

source-checked milestone · 4 Mar 2022 · OpenAI researchers / arXiv

InstructGPT paper submitted

Training language models to follow instructions with human feedback was first submitted to arXiv on 4 March 2022. Diogo Almeida is a listed coauthor.

Why it matters

Provides a dated research reference and the supported authorship credit for Diogo Almeida. It is research context, not evidence of Jev accuracy or an inventor title.

Limits

This is the paper submission date, not a ChatGPT release date. It establishes the research credit; it does not establish a Jev capability, outcome or accuracy result.

Read the decision guides

• Bounded decisions: where Jev fits in an AI workflow
• System One and System Two: a metaphor, not model classes

source-checked milestone · 22 Aug 2022 · Stability AI

Stable Diffusion weights released publicly

Stability AI announced a public release of the Stable Diffusion text-to-image model, following an earlier release to researchers, with the weights under a Creative ML OpenRAIL-M licence that permits commercial and non-commercial use and must accompany any distribution. The release includes an AI-based safety classifier enabled by default whose parameters the user can adjust.

Why it matters

Treat a downloadable image model as a licensing and operations decision, not only a capability one: read the use restrictions that travel with redistribution, decide who accepts them downstream, and budget the hardware and safety-filter settings you would run.

Limits

The announcement describes the release and its stated conditions; it does not establish image quality, legal clearance for any particular use, or suitability for client material. The licence is summarised from the post and the licence text itself should be read before relying on it. Earlier text-to-image and diffusion research existed, and the row does not claim an industry first. No download, local run, quality test or Australian availability check was performed. Prepared with AI assistance.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?

source-checked milestone · 30 Nov 2022 · OpenAI

ChatGPT opens a conversational research release

The dated announcement offers a conversational product for user feedback.

Why it matters

Reconsider whether staff can evaluate a conversational AI interface directly, without first building an application.

Limits

The page does not establish Australian account access or approval to submit client material.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 10 Mar 2023 · Georgi Gerganov (llama.cpp)

llama.cpp runs LLaMA inference and 4-bit quantisation on CPUs without a GPU

A single commit published llama.cpp, a C/C++ implementation of LLaMA inference that converted checkpoints into a quantised ggml file and shipped 4-bit quantisation routines in the same release. The initial code already carried Apple Accelerate and ARM NEON paths, so a model could run on a laptop processor without a discrete GPU.

Why it matters

Treat local inference as a hardware-fit decision: check which runtime and quantisation actually run on the machine you own, and expect the quantisation you choose to trade memory against output fidelity.

Limits

The commit states no benchmarks, throughput figures or supported-hardware list, and none are claimed here. Metal GPU acceleration arrived in June 2023 and GGUF arrived in August 2023; both are later changes, not part of this release. CPU inference and the ggml library predate this commit, so the row makes no industry-first claim. This records a repository commit, not a tagged release; the timeline's evidence labels have no source-commit category, so it is labelled an announcement. No install, run or measurement was performed. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime
• What is AI model quantisation?

source-checked milestone · 14 Mar 2023 · OpenAI

GPT-4 opens text access while image input stays limited

OpenAI announced GPT-4 as a large multimodal model that accepts image and text inputs and emits text, reporting human-level scores on several professional and academic benchmarks. Text input reached ChatGPT Plus users, with API text access through a waitlist, while image input was described as still in limited alpha with a single partner.

Why it matters

Separate what a model can do from what your account can use: before planning on image or long-context input, confirm the model identifier, the context length, and whether that input type is generally available or gated.

Limits

All benchmark comparisons, including the bar-exam result, are OpenAI's claims in its announcement and were not independently reproduced. The row records the launch day only: API access was waitlisted, image input was in limited alpha with one partner, and the 32,768-token variant had limited access, none of which establishes availability to a given account or country. The page has since been revised and the current text is what was read. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Choosing AI: companies, products and models

source-checked milestone · 20 Mar 2023 · OpenAI

ChatGPT cross-user history exposure

OpenAI’s incident report documents a cross-user data exposure and the subsequent fix.

Why it matters

Reassess which client information may enter a shared chat service and how affected material would be identified.

Limits

20 March is the incident day; the report was published on 24 March. This historical event is not evidence of a current leak or that every user was affected.

source-checked milestone · 23 May 2023 · Dettmers, Pagnoni, Holtzman and Zettlemoyer (University of Washington)

QLoRA fine-tunes a 65B model on one 48GB GPU

The QLoRA paper reported fine-tuning a 65-billion-parameter model on a single 48GB GPU by backpropagating through a frozen 4-bit quantised model into low-rank adapters. It introduced a 4-bit NormalFloat data type, double quantisation and paged optimisers. The authors compare the previous requirement for that model, more than 780GB of GPU memory for regular 16-bit fine-tuning, with under 48GB for their method, and report that their best 65B family, Guanaco, reached 99.3% of ChatGPT's performance on the Vicuna benchmark after 24 hours of fine-tuning.

Why it matters

Local control over a model's behaviour is a memory question: for a 65B model this method replaced a multi-GPU requirement with one 48GB card, so weigh adapting a model of that size against needing that workstation GPU and accepting a lossy 4-bit base.

Limits

Every figure here is the authors' own, taken from the version-1 paper; no replication, benchmark run or hardware purchase was performed. 'Single GPU' means one 48GB card, not a consumer device, and consumer-laptop adaptation is not claimed. The Guanaco comparison is a research claim about one benchmark, not a general capability measure. Prepared with AI assistance.

Read the decision guides

• What is AI model quantisation?

source-checked milestone · 13 June 2023 · OpenAI

OpenAI adds function calling to the Chat Completions API

OpenAI announced function calling in its Chat Completions API: developers describe functions to gpt-4-0613 and gpt-3.5-turbo-0613, and the model can choose to return a JSON object of arguments matching the described signature. The announcement presents the feature as a more reliable way to connect model output with external tools and APIs.

Why it matters

Design tool use so the model returns arguments and your own code decides whether to execute them: validate the returned JSON against the signature you supplied, and pin the exact model snapshot before depending on the behaviour.

Limits

The model returns arguments; it does not execute application code, and the reliability claim is the publisher's without a published general benchmark. Function calling is a capability of named API snapshots, not an entitlement on every model, plan or region, and Australian availability was not tested. Tool use existed in other forms before this release, including the plugins pattern the post cites, so the row claims no industry first; the later Model Context Protocol entry is a separate event. The page also carries a 20 July 2023 update note that does not change the announcement date. Prepared with AI assistance.

Read the decision guides

• What may an AI agent read or change? · Shared authority guide
• Is a tool-using agent ready for a bounded pilot?
• Agent harnesses: making AI work inside a controlled workflow

source-checked milestone · 8 July 2023 · Ollama

Ollama packages local models behind a Go server and REST API

llama.cpp already provided local inference; Ollama's v0.0.1 release packaged model downloads, switching and a REST API around that runtime, as a Go inference server that used Metal or CUDA where available. It could be installed as a macOS download, run as a Docker image on port 11434, or built from source; a Windows download was listed as coming soon.

Why it matters

Packaging changes who can run a model: decide whether a project depends on a managed local runtime with its own model store and API, or on a runtime you build and pin yourself.

Limits

These are the release's own claims at the v0.0.1 tag; no install, run or performance measurement was performed. Windows support and the OpenAI-compatible API of February 2024 are later changes and are not attributed to this date. Whether an untagged build circulated before 8 July 2023 was not established. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime

source-checked milestone · 18 July 2023 · Meta

Llama 2 weights for research and commercial evaluation

Meta’s announcement explicitly includes pretrained and conversational model weights.

Why it matters

Reconsider hosted access versus running supplied weights, subject to the exact Llama 2 licence and use policy.

Limits

The announcement is not an unrestricted licence grant or a finding that a client workload is permitted. Exact terms still need review.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 21 Aug 2023 · llama.cpp project

GGUF gives local model files self-describing metadata

llama.cpp merged GGUF, a single-file model format that carries weights together with extensible key/value metadata and the tokeniser information a loader needs. Earlier ggml files already carried model parameters and the vocabulary in a fixed layout that the loader read directly; GGUF added extensible metadata, so a loader could read model details from the file instead of relying on a converter's conventions.

Why it matters

A file that describes itself is easier to move between tools: prefer formats whose metadata travels with the weights, and expect a format change to leave older conversions behind.

Limits

The merge establishes the format change, not its adoption; no usage, download or interoperability figures are claimed. Later extensions such as new quantisation types and sharding are separate work, and no conversion or load was performed here. This records a merged pull request rather than a tagged release; the evidence labels have no merge category, so it is labelled an announcement. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• What is AI model quantisation?

source-checked milestone · 28 Aug 2023 · OpenAI

Enterprise data and identity controls

ChatGPT Enterprise launches with stated business-data and organisation-management protections.

Why it matters

Evaluate an organisation-managed chat environment against client confidentiality and staff-access requirements.

Limits

These product commitments do not establish Australian data residency, contractual suitability or a compliance conclusion.

Read the decision guides

• What may an AI agent read or change? · Shared authority guide

source-checked milestone · 30 Oct 2023 · United States Government

A US executive order sets AI safety reporting duties

The US President signed Executive Order 14110 on the safe, secure and trustworthy development and use of artificial intelligence, directing agencies to set reporting and testing expectations for the most capable models and to address risks across safety, security and civil rights. The order was published in the Federal Register on 1 November 2023 as 88 FR 75191.

Why it matters

Read a national AI instrument as a set of duties with dates and agencies attached: identify which obligations could reach your suppliers or your own deployments before relying on a headline summary, and note that an executive order can be amended or revoked.

Limits

This records the signing, not every duty or agency action that followed, and it is not legal advice. The order's directives were largely addressed to federal agencies; effects on a non-US organisation are indirect and depend on suppliers and on later rules. The instrument was later rescinded by a subsequent administration, so it is a dated governance event and not a description of current US policy. Details are summarised from the Federal Register document and the White House fact sheet. Prepared with AI assistance.

source-checked milestone · 11 Dec 2023 · Mistral AI

Mixtral introduces permissively licensed MoE weights

The release makes a sparse model family available with a named permissive licence.

Why it matters

Evaluate a self-operated mixture-of-experts model under Apache 2.0 as another deployment option.

Limits

The post’s benchmark and inference-efficiency claims are not a verified cost saving for a firm. No Cost secondary type is adopted.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 15 Feb 2024 · Google DeepMind

A limited early access of million-token context

Google opens restricted testing of Gemini 1.5 Pro’s enlarged context window.

Why it matters

Consider a long-document evaluation through the private early access rather than assume the expanded context is generally available.

Limits

Input capacity is not elapsed autonomous working time, perfect recall or universal account access. The snippet’s Capability classification is not retained.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 26 Feb 2024 · Mistral AI

Mistral Large becomes available through Azure

Mistral names Azure as its first distribution partner.

Why it matters

Evaluate the new Azure access path alongside Mistral’s own platform when choosing a provider.

Limits

The announcement does not settle Australian regions, account entitlements or data-handling terms.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 4 Mar 2024 · Anthropic

Claude API becomes generally available

The Claude 3 announcement changes the API access path as well as launching models.

Why it matters

Reconsider direct API evaluation now that developers can sign up for the generally available service.

Limits

Haiku was still forthcoming. A 159-country statement alone does not verify Australian access; model rankings are not the reason for this row.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 18 Apr 2024 · Meta

Meta AI rollout names Australia

The announcement adds named countries to the assistant rollout.

Why it matters

Review staff use of AI inside familiar messaging and social apps as the English-language rollout reaches Australia.

Limits

A rollout is not proof that every account had access that day or that client work is permitted. The page also displays a 3 July update date.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 29 May 2024 · Mistral AI

Codestral introduces non-production licence terms

Mistral introduces a licence family distinct from its Apache 2.0 releases.

Why it matters

Check the non-production terms before treating downloadable coding weights as available for business deployment.

Limits

This records the issuer’s stated boundary, not a legal opinion on any proposed use.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 21 June 2024 · Anthropic

Tool-equipped coding with Claude 3.5 Sonnet

Anthropic describes a coding workflow with tool execution and troubleshooting.

Why it matters

Evaluate a bounded coding task where the model can edit and execute code, rather than only suggest it.

Limits

The statement is conditional on tools and instructions. Benchmark gains and unchanged mid-tier pricing do not create a separate Cost decision.

Read the decision guides

• Longer-running coding agents: when to let them continue
• Web, CLI, IDE, API or messaging: choose an interface
• What longer-running coding does—and does not—prove
• Before staff use an agent on a real repository

source-checked milestone · 23 July 2024 · Meta

Llama outputs may improve other models

The release contains an explicit change to permitted downstream model work.

Why it matters

Reconsider using Llama-generated material to improve another model under the revised terms.

Limits

This is narrower than an unrestricted training-data or licence conclusion; other conditions still apply.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 24 July 2024 · Mistral AI

Large 2 self-deployment has a commercial licence gate

Mistral distinguishes research rights from commercial self-deployment rights.

Why it matters

Budget for a separate commercial licence before choosing Large 2 for commercial self-deployment.

Limits

The same page’s general older-model deprecation statement lacks an exact affected inventory and sunset date; it is not attached as Retirement.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 1 Aug 2024 · European Union

The EU AI Act enters into force

The EU Artificial Intelligence Act entered into force on 1 August 2024, on the twentieth day after its publication in the Official Journal on 12 July 2024, creating a risk-based legal framework for AI providers and deployers in the European Union. Its obligations are staged: most of the regulation applies from 2 August 2026, with prohibitions and AI literacy duties from 2 February 2025.

Why it matters

Treat the entry-into-force date as the start of a compliance clock rather than the start of obligations: map which duty attaches to your role and system, and when, before putting the date into a proposal, contract or client note.

Limits

This is a legal event, not legal advice or a compliance conclusion for any organisation, and Australian obligations do not follow from the EU act. Entry into force, application and each phase carry different dates, and some application dates have since been amended: the Official Journal page retrieved is the act as originally published and notes a current consolidated version dated 27 July 2026, while the European Commission's application timeline was checked on 30 October 2026. The description of the act as the first comprehensive AI legal framework is the European Commission's characterisation, not an independent finding. Prepared with AI assistance.

source-checked milestone · 2 Aug 2024 · DeepSeek

Automatic prefix-cache billing

DeepSeek introduces a distinct billing rule for repeated prompt prefixes.

Why it matters

Recalculate repeated-context workloads using measured cache hits instead of charging every repeated input as a miss.

Limits

Hits are not guaranteed. Historical prices on this page have since changed; no current cost or privacy guarantee is inferred.

source-checked milestone · 22 Aug 2024 · LM Studio

LM Studio 0.3.0 adds document chat, a structured-output API and a network-serving toggle

LM Studio 0.3.0 added chat with your own documents: short documents are inserted in full when they fit the model's context, and longer ones use retrieval to select passages. It also added an OpenAI-like structured-output API that takes a JSON schema, and a toggle to serve the local model to other devices on a network. The desktop app already ran models offline and already exposed a local OpenAI-like server.

Why it matters

Decide how much of a local workflow should live in one application: a frontend that also serves an API and reads your documents replaces several separate tools, at the cost of depending on a closed-source app.

Limits

The release post is the vendor's own description; the features were not verified and nothing was installed. Earlier releases already exposed a local server API, so programmatic access is not new here; the change is scope inside this product. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime

source-checked milestone · 12 Sept 2024 · OpenAI

OpenAI o1 trades latency for reasoning

OpenAI released the first models of its o1 series, which it describes as spending more time computing before responding in order to work through harder science, coding and mathematics problems. ChatGPT Plus and Team users could select the models from launch, with weekly message limits, and API access began for usage tier 5.

Why it matters

Compare a reasoning model against a faster one on your own task before switching: the trade is response time and token cost against accuracy on multi-step problems, and the early access limits decide whether that comparison is possible at your volume.

Limits

OpenAI describes the release as an early access stage and its evaluation figures are the publisher's claims, not an independent assessment; the row asserts no general accuracy improvement. The page records that the two models first released were later retired from the API and that usage limits changed during September 2024, so launch availability is not current availability. No Australian access, cost comparison or Digital Sanctum test is established. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Choosing AI: companies, products and models

source-checked milestone · 22 Oct 2024 · Anthropic

Computer use enters public beta

Anthropic opens computer use to developers as an experimental beta.

Why it matters

Evaluate a supervised screen-operation pilot where the agent can carry out interface steps on the user’s behalf.

Limits

The source calls it error-prone. It does not establish reliable unattended operation or permission to touch client systems.

Read the decision guides

• Agent harnesses: making AI work inside a controlled workflow
• Web, CLI, IDE, API or messaging: choose an interface
• What may an AI agent read or change? · Shared authority guide
• Is a tool-using agent ready for a bounded pilot?

source-checked milestone · 25 Nov 2024 · Anthropic

Model Context Protocol opens a shared integration path

Anthropic releases the MCP specification, SDKs and server implementations.

Why it matters

Evaluate reusable connectors built against an open protocol instead of a separate integration for each AI tool.

Limits

An open protocol does not guarantee interoperability, migration effort or the safety of a particular server.

Read the decision guides

• Agent harnesses: making AI work inside a controlled workflow
• Does MCP make an integration portable? Test the exit
• What may an AI agent read or change? · Shared authority guide
• Is a tool-using agent ready for a bounded pilot?

source-checked milestone · 11 Dec 2024 · Google

Gemini adds delegated Deep Research

The Gemini 2.0 announcement contains a concrete research-assistant feature.

Why it matters

Evaluate delegating an initial research-and-reporting pass, with a person checking its sources and conclusions.

Limits

This row concerns Deep Research, not every Gemini 2.0 prototype. Research demos such as Mariner are not presented as generally available products.

Read the decision guides

• Agent harnesses: making AI work inside a controlled workflow
• Verify Australian access before planning an AI rollout

source-checked milestone · 26 Dec 2024 · DeepSeek

V3 announces a later pricing change

The release announces a future pricing schedule rather than an immediate permanent discount.

Why it matters

Recalculate planned API workloads against the announced post-promotion cache-hit, cache-miss and output rates.

Limits

These are historical announced rates, not current prices. The page’s generic open-source wording is not used as proof of a named weights licence.

source-checked milestone · 20 Jan 2025 · DeepSeek

R1 permits use of weights and outputs under MIT

The release expressly opens downstream use of the model and its outputs.

Why it matters

Evaluate permitted use of R1 weights and outputs for a model-customisation workflow.

Limits

Do not generalise this statement to the separate licences of every distilled model. Listed API prices and benchmarks do not supply extra decision types here.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 5 Feb 2025 · Google DeepMind

Gemini 2.0 Flash API becomes generally available

The page changes the API release stage for 2.0 Flash.

Why it matters

Reassess moving an API evaluation toward production now that Flash has left its experimental access stage.

Limits

The app rollout was the prior week. Pro remained experimental and Flash-Lite was in early access; “everyone” is not a uniform access entitlement.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 25 Mar 2025 · DeepSeek

V3-0324 changes the weights licence to MIT

The page identifies a licence change despite unchanged API usage.

Why it matters

Recheck the permitted deployment and modification path for V3 under its newly stated MIT licence.

Limits

Reasoning and front-end performance claims alone would not pass this gate. The licence statement is the qualifying change.

Read the decision guides

• Open weights, closed weights and the rights your firm needs
• Read the licence attached to the exact weights
• Hosted access versus operating weights: who does what?
• Using model outputs for training and distillation

source-checked milestone · 14 May 2025 · OpenAI

GPT-4.1 reaches paid ChatGPT accounts

The dated release note opens a product access path.

Why it matters

Evaluate GPT-4.1 within an existing paid chat workflow instead of requiring API integration.

Limits

The note describes historical rollout, not current model availability.

Read the decision guides

• Choosing AI: companies, products and models
• AI procurement identity worksheet
• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 22 May 2025 · Anthropic

Claude Code reaches GA across development surfaces

Claude Code leaves research release.

Why it matters

Decide where a coding agent should operate: terminal, editor or background workflow.

Limits

IDE extensions and the GitHub example are beta; product GA does not make every integration GA.

Read the decision guides

• Longer-running coding agents: when to let them continue
• What Claude Code is—and which decisions it owns
• Web, CLI, IDE, API or messaging: choose an interface
• Before staff use an agent on a real repository

source-checked milestone · 22 May 2025 · Anthropic

Claude 4 combines reasoning and tool execution

The model release opens a tool-using reasoning workflow.

Why it matters

Evaluate tasks that require the model to alternate reasoning with tool use.

Limits

This model decision is separate from the Claude Code product’s GA decision announced that day.

Read the decision guides

• Longer-running coding agents: when to let them continue
• What longer-running coding does—and does not—prove
• Before staff use an agent on a real repository

source-checked milestone · 22 July 2025 · Qwen Team

Qwen Code adds a terminal agent tool

The Qwen3-Coder post introduces a concrete CLI application.

Why it matters

Evaluate Qwen’s coding agent in the terminal, with repository and tool access considered separately from model selection.

Limits

The page calls Qwen Code a research-purpose tool. Its model benchmarks do not establish safe repository access or Australian service availability.

Read the decision guides

• Longer-running coding agents: when to let them continue
• Web, CLI, IDE, API or messaging: choose an interface
• Before staff use an agent on a real repository
• What longer-running coding does—and does not—prove

source-checked milestone · 15 Oct 2025 · NVIDIA

NVIDIA opens orders for DGX Spark, a 128GB desktop

NVIDIA's release said DGX Spark could be ordered from 15 October 2025: a desktop built on the GB10 Grace Blackwell platform with 128GB of unified memory and up to 4TB of NVMe storage. The same release states it runs inference on models of up to 200 billion parameters.

Why it matters

For CUDA-dependent local work, weigh a desktop with a large coherent memory pool against a workstation card or a rented GPU; capacity bounds which model fits, and speed depends on bandwidth, compute and the software you run.

Limits

The 200-billion-parameter figure is NVIDIA's own claim and was not tested here. The US$3,999 price is not stated in the text of NVIDIA's newsroom release; The Verge reports it from an infographic published with the announcement, so it is treated as secondary here. Peak-arithmetic figures are not measured throughput. DGX Spark did not raise the absolute memory ceiling, since a 512GB Mac Studio already exceeded it, and no cheaper-than-alternatives claim is made. Partner systems ship their own variants, so this is a platform launch. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Local AI memory budget worksheet: estimate first, then measure
• Local AI total-cost worksheet: compare ownership with hosted use fairly

source-checked milestone · 24 Nov 2025 · Anthropic

Claude Code with Claude Opus 4.5

Anthropic released Opus 4.5 and updated Claude Code that day, describing tools for longer-running agents.

Why it matters

Decide when a longer coding-agent run is worth allowing, and which scope, stop points and checks must come first.

Limits

The announcement does not establish an hours-long run compared with earlier Opus models or every other flagship model. Its quoted 30-minute autonomous coding session is a customer statement, not Anthropic’s own duration finding.

Read the decision guides

• Longer-running coding agents: when to let them continue
• What Claude Code is—and which decisions it owns
• What longer-running coding does—and does not—prove
• Before staff use an agent on a real repository

source-checked milestone · 29 Jan 2026 · OpenAI

ChatGPT announces legacy-model retirement

The announcement gives a dated ChatGPT migration boundary.

Why it matters

Identify chat workflows needing migration before the announced 13 February retirement.

Limits

GPT-5 Instant and Thinking were previously announced for retirement. This is a historical notice, not a live inventory of available models.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label

source-checked milestone · 5 Mar 2026 · OpenAI

GPT-5.4 released with native computer use

OpenAI released GPT-5.4 on 5 March 2026 in ChatGPT, the API and Codex, describing it as the company's first general-purpose model with native computer-use capabilities. The API took the model the same day as gpt-5.4, while ChatGPT and Codex access rolled out gradually.

Why it matters

Evaluate computer use as an API and Codex capability rather than a ChatGPT feature: confirm the exact model identifier, the tools your workflow actually calls and the permissions an agent would hold before piloting any task that operates real software.

Limits

First refers to OpenAI's general-purpose models, not to computer use as an industry first. Same-day API availability, gradual ChatGPT and Codex rollout and plan eligibility are separate facts, and this row records only the release day. Native computer use was announced for the API and Codex, not ChatGPT. Benchmark figures and capability descriptions are OpenAI's claims. No Digital Sanctum API call, task test, latency measurement or Australian eligibility check was performed. Prepared with AI assistance.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Is a tool-using agent ready for a bounded pilot?
• What may an AI agent read or change? · Shared authority guide

source-checked milestone · 12 Mar 2026 · OpenClaw

OpenClaw makes a local model a first-class onboarding choice

Release 2026.3.11 added a first-class Ollama setup path to OpenClaw's onboarding, with Local or Cloud + Local modes, curated model suggestions and cloud-model handling that skips unnecessary local pulls. Before it, a local model was reached by configuring a provider.

Why it matters

When an agent offers local inference during setup, where the model runs becomes a visible choice rather than a configuration detail; decide per workflow whether that model sits on your machine or a provider's.

Limits

This is one release note from a fast-moving project, and the feature was not run here. Local mode applies only when the user selects it: the harness does not run models locally by default, and much documented use is hosted. Adoption of the local path is unmeasured. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime

source-checked milestone · 7 Apr 2026 · Anthropic

Project Glasswing launches with restricted Mythos Preview access

Anthropic announced Project Glasswing, bringing technology and infrastructure organisations together to use Claude Mythos Preview for defensive security work. The launch provided selected organisations with access to the model to help find and fix software vulnerabilities.

Why it matters

Assess how AI-assisted vulnerability discovery changes software-maintenance priorities and the controls needed for model access. Participation in a restricted research programme does not establish access for your organisation.

Limits

This records the 7 April launch, not the later 6 October Cyber Verification Program update now shown on the same page. Anthropic’s capability claims are attributed; no vulnerability count, independent performance measurement or confirmed customer compromise is asserted here. Access at launch was restricted, not general availability; Australian or account eligibility is unresolved. Prepared with AI assistance; no model, exploit or paid API test was performed.

source-checked milestone · 24 Apr 2026 · DeepSeek

DeepSeek V4 makes million-token context the default

DeepSeek released a V4 early access stage with two open-weight models, V4-Pro and V4-Flash, describing one-million-token context as the default across its official services and an API change that keeps the base URL while switching the model name. The announcement also states that its earlier chat and reasoning model names would be retired later in 2026.

Why it matters

Price long-context work against the model you would otherwise use rather than a headline window: check which model name your code calls, whether the context you need is available at the price you assumed, and when the previous model name stops routing.

Limits

Parameter counts, benchmark comparisons and cost claims are DeepSeek's own and were not reproduced or independently benchmarked. The page says the weights are published and links a technical report, but the licence text was not read, so no rights conclusion is drawn. The retirement date is the publisher's plan, not an observed shutdown. The later V4.1 entries are separate, later releases and are not what existed beforehand; the relevant comparison is the model names and context limits this release replaced, which the page states only partly. No Digital Sanctum API call or Australian availability check was performed. Prepared with AI assistance.

source-checked milestone · 28 May 2026 · OpenAI

Two ChatGPT legacy-model sunset dates

One notice sets two product-specific deadlines.

Why it matters

Schedule migration of remaining legacy chat workflows against both deadlines.

Limits

ChatGPT only: the notices say the API was unchanged. OpenAI’s separate Model Release Notes still say 27 June for GPT-4.5, while the ChatGPT release notes give 26 June and confirm removal on that date. The row does not independently audit o3 completion or current model availability.

Read the decision guides

• AI procurement identity worksheet
• Test AI against the work, not the flagship label

source-checked milestone · 21 July 2026 · earendil-works

Pi adds local llama.cpp model management to a coding agent

Pi v0.81.0 added local llama.cpp model management: the agent connects to a llama.cpp router, searches and downloads Hugging Face models, and explicitly loads or unloads them with live progress. Earlier versions already reached local servers through provider configuration in models.json, documented for Ollama, LM Studio and vLLM; the change is lifecycle management of a llama.cpp router rather than the arrival of local inference.

Why it matters

A coding agent that can drive a local runtime keeps inference on your machine when you configure one; check that its model lifecycle is explicit rather than assumed, and do not read it as making the agent's other connections local.

Limits

The release note for v0.81.0 and the repository documentation at that tag are the main sources; the baseline comes from the models documentation at v0.80.0, a different version. Nothing was run or measured. One release note does not show how usable small local models are inside the agent, and it still supports hosted providers. The router address is configurable, so a remote router is possible. Prepared with AI assistance.

Read the decision guides

• Running AI locally: is local, hosted or hybrid operation worthwhile?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime

source-checked milestone · 10 Sept 2026 · DeepSeek

V4.1 Flash adds native visual input and a new API model path

DeepSeek announces V4.1 Flash on 10 September 2026, with native visual input and a documented API migration path. This is the September release, distinct from the later 1017 checkpoint.

Why it matters

Evaluate native text-and-image input through the documented API path; record the exact served model before comparing it with another workflow.

Limits

Architecture, performance and cost comparisons in the announcement are vendor claims, not results reproduced here. Historical API availability does not establish current Australian account eligibility, client-data suitability or the snapshot served by a third-party provider.

reviewed milestone · 15 Sept 2026 · Google

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking

Google announced two variants with different emphases on conversational efficiency and more involved reasoning.

Why it matters

This distinction can help readers separate voice-interface requirements from model behaviour, but the announcement does not prove availability for Australian accounts. Exact API model identifiers, eligible regions and account tiers remain unverified, and no performance claim is adopted here.

Limits

The page says it was updated on 17 September. That later date is revision metadata for the same announcement, not evidence of a second release or, by itself, a substantive claim change.

Read the decision guides

• Verify Australian access before planning an AI rollout
• Test AI against the work, not the flagship label
• Web, CLI, IDE, API or messaging: choose an interface

source-checked milestone · 15 Sept 2026 · TypeSafe

Jev early access announced

TypeSafe announced early access for Jev on 15 September 2026. Its later no-waitlist announcement is recorded as an access revision on this event.

Why it matters

Records TypeSafe’s early-access announcement for Jev. It raises a documented bounded-decision interface to consider, with suitability and current account eligibility still unverified.

Limits

Day-only evidence stays day-only. Historical public access does not establish current Australian account eligibility or an SLA. Jev accuracy in our use is unmeasured because the evaluation was dropped. No speed, saving, reliability or benchmark result is adopted.

Read the decision guides

• Bounded decisions: where Jev fits in an AI workflow
• Noul, Choice and Score: choose the answer shape
• System One and System Two: a metaphor, not model classes
• Where a classifier-shaped model might fit
• Jev’s published price: rate ratios and task bills

source-checked milestone · 28 Sept 2026 · Anomaly / OpenCode

OpenCode v1.18.33: historical MCP OAuth fixed release

OpenCode’s official v1.18.33 release is dated 28 September. Pillar attributes a Windows/WSL MCP OAuth command-injection fix to that version. An AI coding tool is affected; this is an older patch backfill, not a new October patch or evidence of an AI-performed attack.

Why it matters

Inventory the applicable coding-agent version, platform and connected MCP authentication services; assess the developer process’s permissions.

Limits

Pillar’s 19 October header conflicts with its 4 October disclosure chronology, so first-publication dating remains held. The fixed-release date is separate. Patch ancestry is documentary evidence, not a functional security test. No direct Linux web-server or deployment-pipeline compromise is established. This is a local editorial timeline entry, not a native Company event-ledger record; it carries no native acceptance or complete-coverage credit.

Read the decision guides

• PHP2Shell, Tailworm and OpenCode: what changes for website security?

source-checked milestone · 14 Oct 2026 · Nous Research

Hermes Agent installs and manages llama.cpp for local models

Version 0.22.0 gave Hermes Agent a managed local-model path: the agent downloads and manages the llama.cpp runtime, picks a build of the model for the hardware and handles the memory settings that would otherwise be configured by hand. The documentation shipped with that version enabled the desktop Local Models interface for canary builds, required the --local launch flag on other desktop builds, and left choosing and activating a local model as the user's step.

Why it matters

When the agent installs its own inference runtime, setting up a local model becomes a supported path instead of a project of its own; decide whether a runtime you did not build is acceptable in exchange for the setup it removes.

Limits

The release note and the versioned documentation are the vendor's; nothing was installed or measured here. The entry does not show that users run models locally, since hosted providers remain supported. The same documentation distinguishes a runtime that may already be bundled from the managed-tool install path, so what exists on a given machine varies. The release text dates itself 13 October 2026 while the platform published it on 14 October. Prepared with AI assistance.

Read the decision guides

• Hermes Agent: does it fit a workflow you can inspect and maintain?
• Ollama, llama.cpp, vLLM or LM Studio: choosing a local AI runtime

source-checked milestone · 17 Oct 2026 · DeepSeek

Flash 1017 arrives and Search/Fetch API access opens

On 17 October 2026 DeepSeek announces the Flash 1017 update and broader access to its Search and Fetch interfaces. Benchmark-maintainer coverage and first-hand local inference reports provide distinct evidence of the attention around the release.

Why it matters

Test the named 1017 checkpoint and supported harness/tool integration on a bounded workload; keep publisher claims, external assessments and your own measured outcomes separate.

Limits

Reported gains are attributed to their evaluator and settings. Search and Fetch are separate APIs, not intrinsic browsing authority. No industry consensus, universal superiority, current Australian entitlement or 1017 provenance for our unpinned CommandCode benchmark is established.

source-checked milestone · 21 Oct 2026 · PHP / The PHP Foundation

PHP2Shell disclosure and AI-assisted security research

The PHP Foundation’s 21 October account describes two critical engine vulnerabilities, AI-assisted security work and human investigation/patch decisions. Key advisory timestamps cross into 22 October AEDT. Relevant deployment conditions and vendor fixes must be checked.

Why it matters

Review PHP runtime and vendor patch evidence alongside application updates. AI-assisted research is the connection; AI attacker use is not established.

Limits

The source labels its calendar date as 21 October; its precise publication timezone is not inferred. The two advisories’ UTC timestamps fall on 22 October in Melbourne. Paths/configurations differ, the 23 October session follow-up is separate, and attempted exploitation is not proof of universal compromise. This is a local editorial timeline entry, not a native Company event-ledger record; it carries no native acceptance or complete-coverage credit.

Read the decision guides

• PHP2Shell, Tailworm and OpenCode: what changes for website security?

source-checked milestone · 26 Oct 2026 · OpenAI

GPT-6.1 Terra API release

OpenAI released GPT-6.1 Terra in its API. The publisher describes the model as supporting visual reasoning, computer use and writing; its model documentation lists text and image inputs with text output.

Why it matters

Evaluate gpt-6.1-terra for image-and-text workflows. Use the Responses API when tools are needed, and test the tasks and permissions your workflow actually requires.

Limits

The date records the API changelog’s release day; no time or timezone is supplied. Capability descriptions are OpenAI’s claims, not Digital Sanctum benchmark results. Work and Codex access has separate plan, workspace and setup conditions; API release does not establish universal account or Australian eligibility. Prepared with AI assistance; no paid API or functional capability test was performed.