Part of IT Trends & Reviews — what actually shipped, quarter by quarter.
1. Introduction: The Open-Source AI Summer
The third quarter of 2023 is, above all, the quarter when large language models escape the walled garden. For the previous eighteen months, the most capable AI systems sit behind paywalls, API rate limits, and opaque terms of service. In practice, a handful of American technology companies control them. Then, on July 18, Meta releases Llama 2 in full partnership with Microsoft. Notably, weights are included, commercial use is permitted, and a research paper documents every design choice. As a result, the landscape changes immediately and irreversibly.
Moreover, the implications are not merely academic. Within days of the Llama 2 release, engineers at companies across every industry are running 7-billion and 13-billion parameter models on their own servers. They fine-tune them on proprietary data and deploy conversational interfaces without sending a single token to a third-party API. The democratisation of inference that open-source advocates have demanded since ChatGPT’s November 2022 shock is now, technically, possible.
However, Llama 2 is only the opening act of a summer defined by competitive releases. On August 24, Meta follows with Code Llama — a family of code-specialised variants that challenge GitHub Copilot on its own territory. On September 27, a six-month-old Parisian startup named Mistral AI drops its first model, Mistral 7B. Notably, distribution runs over a BitTorrent magnet link posted on X (formerly Twitter). The company claims it outperforms Llama 2 13B on standard benchmarks despite having only 7.3 billion parameters. As a result, the release instantly transforms Mistral from an unknown entity into the most-discussed AI startup in Europe.
1.1 The proprietary frontier, regulation, and the GPU bottleneck
Meanwhile, the proprietary frontier continues to advance. Meanwhile, Anthropic releases Claude 2 in July, doubling the context window to 100,000 tokens and matching or exceeding GPT-4 on several legal and technical benchmarks. OpenAI ships ChatGPT Enterprise on August 28, establishing clear commercial pricing and data governance terms that finally allow regulated industries to consider LLM adoption. In addition, Amazon’s Bedrock managed-AI service moves toward general availability with Anthropic, Meta, and Stability AI models all available through a unified cloud API.
Meanwhile, the political dimension of AI also crystallises this quarter. Specifically, the European Parliament voted to approve the EU Artificial Intelligence Act in June. Trilogues between the Parliament, Council, and Commission are now in progress to finalise a text. That text introduces risk-based obligations for AI systems and transparency requirements for generative AI. In addition, fines reach up to 6 percent of global annual revenue for the most serious violations. Consequently, AI regulation, long theoretical, is now a matter of engineering compliance.
The GPU supply chain remains the physical bottleneck behind all of this ambition. Nvidia’s H100 Tensor Core GPUs are the preferred hardware for both training and inference. As a result, they carry wait lists of six months or longer at most cloud providers. Surging LLM demand and constrained supply drive Nvidia’s market capitalisation above one trillion dollars for the first time in late May 2023. Moreover, that level holds through Q3 as the company reports record earnings. The GPU shortage forces many teams to optimise aggressively. That pressure accelerates the adoption of quantisation techniques through tools like llama.cpp that allow models to run on consumer hardware.
2. GitHub Deep Dive – Key Repositories of Q3 2023
GitHub activity in Q3 2023 reflects the defining tension of the quarter. Specifically, that tension is the race between open-weight LLMs and the tooling required to run them efficiently without a data centre. In particular, five repositories stand out for their technical significance and community momentum during July, August, and September 2023.
2.1 Meta Llama 2 — meta-llama/llama
Llama 2 — Open-Weight Foundation Models for Research and Commerce
Llama 2 arrives as three model sizes: 6.7B, 13B, and 69B parameters. In addition, each is available in both foundation and instruction-tuned (chat) variants. Overall, the architecture is a decoder-only transformer closely following Llama 1, with two significant improvements. Meanwhile, the training dataset grows from 1.4 to 2 trillion tokens. In addition, the context window doubles from 2,048 to 4,096 tokens. For the chat variants, Meta applies reinforcement learning from human feedback (RLHF) using a proprietary dataset of approximately 27,540 prompt-response pairs.
The safety training introduces a “Ghost Attention” technique that maintains system-prompt compliance across multi-turn dialogues, addressing a persistent weakness of instruction-tuned models. Overall, Meta trains the models on an estimated 21,000 petaFLOP-days of compute, using approximately 4,000 A100 GPUs over several months. This release is a genuine milestone. For the first time, organisations with fewer than 700 million monthly active users can download and deploy commercially competitive LLM weights. Importantly, no API dependency or per-token billing is involved. Its GitHub repository accumulates tens of thousands of stars within the first 48 hours. Hugging Face Hub mirrors receive hundreds of thousands of downloads in the first week.
Nevertheless, the licensing nuance matters. Llama 2 is not strictly open-source by the Open Source Initiative’s definition, because the acceptable use policy restricts certain applications including training other LLMs. It is nevertheless functionally open enough to ignite an ecosystem. Indeed, within weeks, derivative fine-tunes appear for medicine, law, finance, and coding. In the language of Simon Willison, the Llama 2 release is the LLM equivalent of Stable Diffusion’s open-weight release in August 2022. In other words, it is the moment the technology proliferates beyond any single company’s control.
2.2 Code Llama — facebookresearch/codellama
Code Llama — Specialised LLM for Code Generation and Completion
Code Llama is a family of code-specialised models derived from Llama 2 through a two-stage training procedure. Starting from the Llama 2 foundation models, Meta AI trains an additional 500 billion tokens of code-focused data, producing the Code Llama base. In addition, a further 20 billion tokens of long-context data extend the context window to up to 100,000 tokens for the “Code Llama — Code” variant.
The 7B, 13B, and 34B sizes each ship with three sub-variants. Specifically, these are the base code model, a Python-specialised version (trained on an additional 100B tokens of Python code), and an instruction-following instruct version. On the HumanEval benchmark, Code Llama 34B scores 53.7 percent pass@1, competitive with GPT-3.5 and considerably ahead of the general Llama 2 70B. The 100K-token context window is practically significant: engineers can feed entire codebases, not just individual files, into the model for context-aware refactoring or documentation generation. As a result, Code Llama becomes within weeks the default open-weight alternative to GitHub Copilot for teams unwilling or unable to use cloud-hosted code AI tools.
2.3 Mistral 7B — mistralai/mistral-src
Mistral 7B — The European Challenger That Rewrites Efficiency Expectations
Mistral 7B is the debut model from Mistral AI, a Paris-based startup founded in April 2023. Notably, its three founders are former employees of Google DeepMind and Meta: Arthur Mensch, Guillaume Lample, and Timothée Lacroix. Moreover, the company raises €105 million in June 2023, achieving a valuation of approximately €240 million before shipping a single product. When Mistral 7B arrives on September 27 via a torrent magnet link — an intentionally unconventional distribution channel — the benchmark numbers generate immediate excitement.
The model uses Grouped-Query Attention (GQA) for faster inference and a sliding window attention mechanism to handle sequences up to 8,192 tokens efficiently. Mistral AI claims that Mistral 7B outperforms Llama 2 13B on all tested benchmarks, including MMLU, HellaSwag, Winogrande, and Arc. In fact, independent evaluations confirm the claim, despite the model having roughly half the parameter count. The Apache 2.0 license is genuinely permissive, without the commercial-use restrictions of the Llama 2 community license. In practice, that makes Mistral 7B immediately attractive to commercial builders. Above all, the model’s efficiency demonstrates that architectural innovation, not just raw scale, remains a competitive lever in the LLM space.
2.4 llama.cpp — ggerganov/llama.cpp
llama.cpp — Running LLMs on Consumer Hardware
Created by Georgi Gerganov and first released in March 2023, llama.cpp becomes the infrastructure layer for the open-source LLM summer. The project re-implements LLM inference in pure C/C++, with no external dependencies beyond a standard C compiler. Models can therefore run on CPUs, including laptops and Raspberry Pis, without requiring Nvidia GPUs.
Through aggressive quantisation (4-bit and 8-bit integer types), llama.cpp reduces the memory footprint of a 7B model. Specifically, it drops from roughly 14 GB in float16 to approximately 3.5–4 GB in Q4_K_M format. As a result, that fits comfortably in the RAM of a modern MacBook. When Llama 2 ships in July, the llama.cpp team adds support within hours. The GGUF file format, introduced this quarter as a replacement for the earlier GGML format, becomes the de facto standard for distributing quantised open-weight models. Moreover, that repository surpasses 40,000 GitHub stars during Q3, making it one of the fastest-growing ML infrastructure projects of 2023. The practical consequence is enormous. Indeed, running a GPT-3.5-class model locally, on a developer laptop, without internet access, without API costs, and without data leaving the machine, is now routine.
2.5 LangChain — langchain-ai/langchain
LangChain 0.0.x — The Orchestration Layer for LLM Applications
LangChain, created by Harrison Chase and released in late 2022, continues its extraordinary growth trajectory in Q3 2023 with over 60,000 GitHub stars. Specifically, the project provides abstractions for chaining LLM calls, embedding retrieval-augmented generation (RAG) pipelines, and building AI agents that use external tools.
During Q3 2023, the team introduces LangChain Expression Language (LCEL), a declarative composition syntax. In practice, it allows developers to combine retrieval, prompting, model calls, and output parsing into concise, readable pipeline definitions. This shift toward LCEL marks an architectural maturation. Rather than write imperative Python scripts that string together individual LangChain components, teams can declare data-flow pipelines that are more testable and composable. The LangChain ecosystem — including LangSmith for observability and LangServe for deployment — grows into a platform. The library’s broad Llama 2 and local model integrations make it the default scaffolding for many teams. In particular, they build RAG-based enterprise search, document question-answering, and conversational agents with it in Q3 2023.
3. Big Tech & Industry Breakthroughs
Q3 2023 is one of the most active quarters in recent memory for consequential product and strategic moves. Both the established technology giants and a new wave of AI-first companies contribute. The table below captures the most significant events; the subsections that follow explore each in detail.
| Company | Event | Date | Significance |
|---|---|---|---|
| Meta | Llama 2 released with Microsoft | Jul 18, 2023 | First commercially viable open-weight LLM; triggers open-source AI summer |
| Anthropic | Claude 2 launched | Jul 11, 2023 | 100K-token context window; strong legal/coding benchmarks; Claude.ai consumer access |
| OpenAI | ChatGPT Enterprise shipped | Aug 28, 2023 | Enterprise data governance terms; no training on customer data; SSO and admin controls |
| OpenAI | Code Interpreter renamed Advanced Data Analysis | Aug 2023 | Available to all Plus users; Python sandbox for data analysis and visualisation |
| Meta | Code Llama released | Aug 24, 2023 | Open-weight code AI; 100K context window; challenges GitHub Copilot |
| AWS / Amazon | Amazon Bedrock preview expands | Q3 2023 | Managed foundation model API; Anthropic, Meta, Stability AI all available |
| Microsoft | Microsoft 365 Copilot GA date announced | Sep 21, 2023 | November 1 GA for enterprise; $30/user/month; AI in Word, Excel, Teams |
| Mistral AI | Mistral 7B released | Sep 27, 2023 | Apache 2.0; outperforms Llama 2 13B; European AI sovereignty signal |
| xAI | Company incorporated by Elon Musk | Jul 2023 | New frontier-AI lab; direct competition signal against OpenAI |
| Apple | iPhone 15 and A17 Pro chip announced | Sep 12, 2023 | First 3-nm mobile chip; Neural Engine advances; USB-C transition |
| Stability AI | Stable Diffusion XL 1.0 released | Jul 26, 2023 | Higher-resolution, improved image quality; open-weight image generation |
| TII (UAE) | Falcon 180B released | Sep 6, 2023 | 180B open-weight model trained on 3.5T tokens; tops open LLM leaderboard |
| EU | AI Act trilogues in progress | Q3 2023 | First comprehensive AI regulation approaching final text |
3.1 Meta and the Open-Weight Bet
Meta’s decision to release Llama 2 under a permissive commercial license is a deliberate strategic move. CEO Mark Zuckerberg articulates the logic publicly. Open AI models benefit Meta because they commoditise the model layer, where Meta has less structural advantage than infrastructure players like Google and Microsoft. At the same time, open models also allow Meta to compete on products and distribution. Equally strategic is the Microsoft partnership. By making Llama 2 the first open-weight LLM available directly in the Azure AI model catalogue, Meta ensures enterprise distribution. Consequently, no cloud sales force of its own is required.
Meta’s Llama 2 training numbers are significant. The foundational models train on approximately 2 trillion tokens — 40 percent more than Llama 1 — with upsampling from sources considered trustworthy. The chat variants additionally incorporate RLHF using over 1.4 million Meta-curated human preference examples across seven supplementary datasets. This produces a model that, on the MT-Bench and Vicuna evaluation frameworks, scores competitively with GPT-3.5 on many conversational tasks at the 70B parameter scale.
Free ebook
Free AI Video, Generated Locally
Working scripts and measured benchmarks. Free.
No spam. Unsubscribe at any time.
3.2 Anthropic’s Claude 2 and the 100K Context Window
Anthropic releases Claude 2 on July 11, 2023, available both via the Anthropic API and through the newly launched Claude.ai consumer interface. Notably, the headline feature is a 100,000-token context window — roughly 75,000 words, or the equivalent of a short novel. This capability is immediately significant for legal document review, financial analysis, and codebase understanding. Indeed, such tasks need the full context of a long document to fit inside a single model call.
On benchmark comparisons, Claude 2 scores 76.5 percent on the Codex HumanEval coding benchmark and 88.0 percent on the Bar Exam passage. Overall, both figures improve substantially on its predecessor. In addition, the model is designed with constitutional AI principles, and Anthropic publishes additional research on its alignment approach. Claude 2’s release reinforces Anthropic’s positioning as the safety-focused alternative to OpenAI. Crucially, that narrative resonates with enterprise customers increasingly concerned about AI governance and liability.
3.3 OpenAI: ChatGPT Enterprise and the Business Market
OpenAI launches ChatGPT Enterprise on August 28, 2023, addressing the most frequently cited concern from corporate IT departments: data privacy. The enterprise tier guarantees that customer inputs and outputs are not used to train OpenAI models, and it provides a Service-Level Agreement. In addition, admins can manage user accounts through SSO integration, and the GPT-4 context window extends to 32,000 tokens by default. However, pricing is not publicly disclosed but is structured per-user at enterprise scale.
Also in August 2023, OpenAI renames the Code Interpreter plugin to Advanced Data Analysis. The plugin provides ChatGPT Plus subscribers with a sandboxed Python runtime for data analysis, file uploads, and chart generation. Renaming it signals a repositioning toward business audiences. It also makes the feature available to all Plus subscribers, not just those who have explicitly enabled the plugin. In practice, the ability to upload a CSV, ask natural-language questions, and receive automatically generated charts and statistical summaries represents a genuinely novel capability for non-technical analysts.
3.4 Amazon Bedrock and the Managed LLM Race
AWS continues expanding Amazon Bedrock through Q3 2023. Bedrock is a fully managed service that allows enterprises to access foundation models from multiple providers. Specifically, the catalogue covers Anthropic’s Claude, Meta’s Llama 2, Stability AI’s image generation models, and Amazon’s own Titan models. Moreover, access runs through a unified API with standard AWS authentication, logging, and data residency controls. The service removes the integration overhead of managing multiple AI vendor relationships and provides AWS’s enterprise customers with a familiar governance wrapper around LLM capabilities. The managed-AI model is significant. Enterprises can adopt AI services under existing AWS contracts and compliance frameworks rather than negotiating new agreements with individual AI startups.
3.5 Microsoft: The Copilot Vision Sharpens
Microsoft announces on September 21, 2023 that Microsoft 365 Copilot reaches general availability on November 1, 2023. Specifically, the price is $30 per user per month as an add-on to existing Microsoft 365 Business or Enterprise subscriptions. The product integrates GPT-4 into Word, Excel, PowerPoint, Outlook, and Teams. In practice, users can generate document drafts, summarise meeting transcripts, create formulas, and receive answers over corporate document libraries. The $30 per-user price point at scale represents one of the largest AI product introductions in terms of potential revenue impact. For example, a Fortune 500 company with 50,000 knowledge workers represents $18 million per year in Copilot fees alone.
Microsoft’s Azure AI Studio, meanwhile, continues to expand model availability. The Azure marketplace adds Llama 2 models immediately upon Meta’s release, positioning Azure as the enterprise distribution channel for open-weight AI. The combination of OpenAI exclusivity, Llama 2 marketplace access, and the Copilot product suite makes Microsoft arguably the most comprehensively positioned company in enterprise AI during Q3 2023.
3.6 Falcon 180B: The UAE Enters the Frontier
On September 6, 2023, the Technology Innovation Institute (TII) in Abu Dhabi releases Falcon 180B. Specifically, it is a 180-billion parameter open-weight model trained on 3.5 trillion tokens of the RefinedWeb dataset. At release, Falcon 180B tops the Hugging Face Open LLM Leaderboard, outscoring both Llama 2 70B and LLaMA 65B on standard benchmarks. In addition, the model is released under the Falcon 180B TII License for research and commercial use. This Falcon 180B release is a geopolitical signal as much as a technical one. In short, frontier-scale open-weight model training is no longer the exclusive domain of US hyperscalers. The UAE government has invested heavily in AI infrastructure, and TII’s Falcon series demonstrates that those investments yield globally competitive results.
3.7 The EU AI Act: Regulation Reaches the Finish Line
By Q3 2023, the EU Artificial Intelligence Act enters trilogue, following the Parliament’s vote to adopt its negotiating position in June 2023. In other words, trilogue is the negotiation between the European Parliament, Council, and Commission. The draft text introduces a risk-based classification with four tiers. One is unacceptable risk (banned, including real-time biometric surveillance in public spaces and social scoring). Next comes high risk, requiring conformity assessments for systems in critical infrastructure, hiring, and law enforcement. Then comes limited risk (requiring transparency disclosures, such as chatbots identifying themselves as AI), and finally minimal risk (unregulated).
For general-purpose AI models above a compute threshold, the Act introduces obligations around training data documentation, copyright compliance, and adversarial testing. In addition, fines reach 7 percent of global annual turnover for the most serious violations. Consequently, technology companies with global operations are actively monitoring the process and beginning AI governance programmes in preparation for compliance.
4. AI & Technology Impact
4.1 Timeline of Key AI Events in Q3 2023
Anthropic Claude 2 Released
Anthropic releases Claude 2 via API and claude.ai. The 100,000-token context window sets a new standard for long-document processing. Scores 76.5% on HumanEval and 88.0% on the Bar Exam passage benchmark.
Meta Llama 2 Released in Partnership with Microsoft
Three model sizes (7B, 13B, 70B), foundation and chat variants, commercial license permitted for most organisations. Trained on 2 trillion tokens with a 4,096-token context window. Triggers a cascade of derivative fine-tunes and downstream tools within 48 hours.
Elon Musk Incorporates xAI
Musk founds xAI Corp as a direct competitor to OpenAI, recruiting researchers from DeepMind and other frontier AI labs. The company’s stated mission is “to understand the true nature of the universe.” No product ships this quarter. However, the competitive signal is noted industry-wide.
Stable Diffusion XL 1.0 Released
Stability AI releases SDXL 1.0, featuring a 6.6B parameter base model, improved composition and photorealism at 1024×1024 resolution, and a refiner model for detail enhancement. Released under a modified CreativeML Open Rail+M license.
OpenAI Renames Code Interpreter to Advanced Data Analysis
This sandboxed Python environment is available to all ChatGPT Plus subscribers. Users can upload files, run code, generate charts, and perform statistical analysis through natural language. The rename signals a shift toward business user positioning.
4.2 Q3 2023 timeline: late August and September
Code Llama Released by Meta
Sizes 7B, 13B, 34B — base, Python-specialised, and instruct variants. Up to 100K token context window. Code Llama 34B scores 53.7% on HumanEval pass@1, competitive with GPT-3.5 Turbo on code generation tasks.
ChatGPT Enterprise Launched by OpenAI
Enterprise data governance, no training on customer data, SSO integration, admin controls, and a 32K token context window by default. Represents OpenAI’s entry into regulated enterprise sales at scale.
Falcon 180B Released by TII (UAE)
Technology Innovation Institute releases Falcon 180B, a 180B parameter open-weight model trained on 3.5 trillion tokens. At release, it tops the Hugging Face Open LLM Leaderboard, demonstrating that frontier-scale open-weight training is no longer exclusive to US hyperscalers.
Apple iPhone 15 and A17 Pro Announced
Apple announces the iPhone 15 series. The A17 Pro chip is the industry’s first 3-nanometre mobile processor. It delivers a 20 percent CPU performance improvement and a significantly more capable Neural Engine for on-device AI inference.
Microsoft 365 Copilot GA Date Set for November 1
Microsoft announces November 1 general availability at $30/user/month. GPT-4 integrated into Word, Excel, PowerPoint, Outlook, and Teams for all qualifying enterprise subscriptions.
Mistral 7B Released Under Apache 2.0
Mistral AI releases its first model via a torrent magnet link on X. The 7.3B parameter model uses Grouped-Query Attention and sliding window attention, achieving benchmark results above Llama 2 13B. Its Apache 2.0 license makes it immediately attractive with no commercial restrictions.
4.3 The GPU Supply Chain and the Inference Efficiency Race
The physical constraint of AI in Q3 2023 is compute availability. Nvidia’s H100 80GB SXM5 GPU, the preferred card for both large-model training and high-throughput LLM inference, remains severely supply-constrained through the quarter. As a result, cloud providers report allocation wait times of six months or more for dedicated H100 instances. In addition, spot pricing on AWS and Azure for H100 nodes frequently exceeds $10 per GPU-hour. The scarcity is structural: TSMC’s advanced packaging capacity for the H100’s HBM3 memory stacks is a bottleneck that cannot be resolved on short notice.
Consequently, this supply pressure accelerates two important technical trends. First, quantisation becomes mainstream. The GGUF format standardised by llama.cpp enables 4-bit and 8-bit quantised inference on CPU hardware. As a result, model size falls by 60–75 percent with acceptable accuracy loss on most tasks. Second, inference optimisation libraries emerge as essential infrastructure. Tools such as vLLM (developed at UC Berkeley), TensorRT-LLM (from Nvidia), and CTranslate2 serve LLMs in production without wasting expensive GPU capacity. vLLM’s PagedAttention algorithm demonstrates throughput improvements of up to 24x over standard Hugging Face Transformers serving, fundamentally changing the economics of LLM deployment.
China’s technology sector faces a different GPU challenge. US export controls implemented in October 2022 and expanded in October 2023 restrict the export of Nvidia’s most capable chips to China. Specifically, the restricted list includes the H100 and A100. Chinese AI labs such as Baidu, Alibaba DAMO Academy, and Zhipu AI are actively developing models on domestic GPU alternatives. In particular, these include Huawei’s Ascend 910B and Cambricon processors. As a result, the regulatory situation forces accelerated Chinese investment in domestic semiconductor design and manufacturing, a multi-year strategic shift.
4.4 The Multimodal Frontier
Text-only LLMs are no longer the frontier by Q3 2023. Google’s PaLM 2, released in May 2023, underpins a multimodal extension called Gemini (still in development this quarter). GPT-4 Vision is the multimodal variant of GPT-4, capable of analysing images alongside text. However, it is available in limited preview to select research partners in Q3, with broader release planned for Q4. The combination of visual understanding with language reasoning enables document parsing, diagram analysis, and medical image interpretation that purely text-based models cannot support. Ultimately, the multimodal direction is clear. In short, future frontier models process text, images, audio, and structured data interchangeably, collapsing the distinction between separate specialised AI systems into unified general-purpose inference engines.
Key Insight: The Open vs. Closed Divide Crystallises
Q3 2023 makes the open-versus-closed LLM question concrete and consequential. Llama 2, Code Llama, Mistral 7B, Falcon 180B, and Stable Diffusion XL 1.0 all ship with weights accessible to anyone. Against them stand GPT-4, Claude 2, and PaLM 2 — accessible only via API, with no visibility into architecture or training data. Moreover, the open models are rapidly closing the benchmark gap. Furthermore, organisations in healthcare, finance, and legal services can run models entirely within their own infrastructure, without sending patient records or financial data to third-party APIs. For them that ability is not just a preference but a regulatory requirement. The open-weight ecosystem is not yet matching GPT-4 on the most demanding tasks. However, the trajectory and pace of improvement suggest the gap narrows significantly through 2024.
5. Key Voices & Thought Leaders
Q3 2023 is unusually rich in influential technical content. The open-source LLM summer generates an enormous body of tutorial writing, benchmark analysis, and architectural commentary. Notably, five voices are particularly influential during this quarter.
5.1 Simon Willison — simonwillison.net
Simon Willison — The Definitive LLM Experimentalist
Simon Willison — co-creator of the Django web framework — publishes an extraordinary volume of technically precise, reproducible LLM experiments throughout Q3 2023. He posts on running Llama 2 locally and on comparing quantised model quality across formats. Similarly, his write-ups on building the llm command-line tool for interacting with local and remote models reach tens of thousands of engineers. He frames the Llama 1 moment as analogous to Stable Diffusion’s democratisation of image generation. As a result, that piece, published in March 2023 and immediately relevant again with Llama 2’s arrival, is widely cited. Willison’s distinctive contribution is making cutting-edge LLM research reproducible and accessible to practising engineers without ML backgrounds. He does it through detailed step-by-step blog posts with exact commands and environment specifications.
5.2 Andrej Karpathy — YouTube and X (formerly Twitter)
Andrej Karpathy — Neural Networks from First Principles
Former Tesla Director of AI and founding OpenAI member Andrej Karpathy rejoins OpenAI in February 2023. He remains one of the most trusted voices explaining transformer architectures and LLM internals to a broad technical audience. His “Neural Networks: Zero to Hero” YouTube series includes the “Let’s build GPT from scratch” video, which accumulates millions of views through 2023. The series gives tens of thousands of engineers a ground-level understanding of how the attention mechanism, tokenisation, and training loops actually work. In Q3 2023, Karpathy’s commentary on the Llama 2 release and Code Llama benchmarks on X is widely circulated. It provides context on what the benchmark numbers mean in practice, and where they mislead. His directness about the limitations of benchmark gaming makes his assessments particularly valuable in a quarter full of competing model claims.
5.3 Yann LeCun — Meta Chief AI Scientist
Yann LeCun — The Open-Source AI Philosopher
Meta’s Chief AI Scientist uses the Llama 2 release as the occasion for sustained advocacy. His position is that open-source AI is categorically safer and more democratically desirable than closed proprietary systems. LeCun argues throughout Q3 2023 that concentrating frontier AI capability inside a small number of proprietary systems is dangerous. It creates single points of failure and regulatory capture. Open models, by contrast, allow distributed scrutiny, fine-tuning for safety, and academic oversight. He also pushes back on the framing of current LLMs as approaching artificial general intelligence. Autoregressive language models, he argues, have fundamental architectural limitations in reasoning and world-model construction that scale alone does not overcome. His position sets the intellectual frame for the open-source AI movement’s expansion in 2023 and 2024.
5.4 Arthur Mensch — Mistral AI CEO
Arthur Mensch — Building the European AI Frontier
Arthur Mensch emerges as the most prominent European AI founder of 2023. Mistral 7B’s September 27 release triggers an industry-wide reassessment of what a small team with focused architectural research can achieve. The decision to release under Apache 2.0, without the commercial-use restrictions of Meta’s Llama license, is a deliberate statement about Mistral’s philosophy. That philosophy holds that models should be genuinely free, not “open-washed”. Mensch’s interviews and public commentary throughout Q3 emphasise that European AI independence is technically achievable and commercially viable. That directly challenges the narrative that frontier AI requires US hyperscaler resources. The model’s benchmark performance, beating Llama 2 13B at 7B parameters, validates the claim. It establishes Mistral as a credible alternative voice to the OpenAI/Google/Meta triad.
5.5 Sebastian Raschka — ML Research Explainer
Sebastian Raschka — LLM Architecture Deep Dives
Sebastian Raschka publishes some of the most technically rigorous LLM architecture analyses of Q3 2023. His newsletter pieces compare Llama 2 and Mistral 7B at the architectural level, explaining grouped-query attention, rotary positional embeddings, and the practical implications of different normalisation choices. They are essential reading for engineers who want to understand why Mistral 7B achieves its efficiency gains. Raschka’s explainers bridge the gap between academic papers and engineering practitioners who think in code. His work this quarter covers parameter-efficient fine-tuning methods including LoRA and QLoRA. It gives thousands of teams the understanding needed to customise open-weight models on domain-specific data without access to large GPU clusters.
6. Trend Synthesis
The events of Q3 2023 collectively signal a structural shift in how AI capability is distributed, governed, and consumed. This quarter, more than any single previous three-month period, rewrites the assumptions about who controls AI and on whose terms.
The first signal is the collapse of the closed-model premium for many use cases. Three months earlier, the best available LLMs all required API access and sent data to third-party servers. By the end of September 2023, an engineer with a laptop can run a competent open-weight model entirely locally at no marginal cost. The Mistral 7B result matches or exceeds Llama 2 13B on standard benchmarks at roughly half the size. That demonstrates that the efficiency frontier of open models advances faster than raw parameter scaling. The practical implication is that the gap between open and proprietary models is closing faster than most observers predicted six months earlier.
The second signal is the emergence of the enterprise AI stack. ChatGPT Enterprise, Microsoft 365 Copilot’s announced pricing, and Amazon Bedrock’s managed API all point toward the same destination. AI becomes a line item in enterprise software budgets. Such spending is governed by the same compliance, security, and procurement frameworks as any other SaaS product. The rough price signal from Microsoft — $30 per knowledge-worker per month — establishes a market rate. Cloud providers emerge as the dominant distribution channel, and not because they train the best models. Enterprises already have procurement relationships, compliance certifications, and billing integrations with AWS, Azure, and GCP. The AI model layer commoditises relative to the distribution and integration layer.
6.1 Geographic diversification and the compliance axis
The third signal is the geographic diversification of AI power. Mistral AI’s September debut establishes that European researchers, without hyperscaler resources, can reach competitive performance through architectural innovation. The Technology Innovation Institute’s Falcon 180B release from Abu Dhabi demonstrates that frontier-scale model training no longer requires Silicon Valley affiliation. China’s AI labs, operating under GPU export controls, are pursuing domestic hardware alternatives and open-weight model strategies. The AI landscape that looked US-dominated in early 2023 is, by the end of Q3, genuinely multipolar at the open-weight tier.
A fourth signal is the acceleration of the regulatory-compliance axis. The EU AI Act’s approaching final text forces global technology companies to begin concrete compliance planning, not just lobbying. Risk classification systems, model documentation requirements, and transparency obligations are not yet in force. However, the architectural decisions made in products built in 2023 determine how easily they satisfy requirements when the law takes effect. AI governance — model cards, training data documentation, bias evaluations, human oversight mechanisms — is transitioning from voluntary best practice to pending legal requirement.
6.2 On-device inference and what the quarter settles
The fifth and most consequential signal is the normalisation of on-device and local inference. Three developments point toward a near future in which capable AI inference happens not in a data centre but on the device. These are llama.cpp’s aggressive CPU quantisation, the GGUF format standard, and Apple’s A17 Pro chip’s improved Neural Engine. That device may be the one in a user’s pocket, or a laptop without a network connection. The implications for privacy, latency, and cost are profound. Conversational AI that never transmits data to a server eliminates entire categories of data residency and privacy concern. It also reduces the incremental cost of a model interaction to nearly zero. The inference efficiency race accelerates through Q3 2023. It is, ultimately, a race toward AI that is faster, cheaper, and more private than what the API-first model permits.
Key Insight: The Open-Source AI Summer Defines the Next Decade
Q3 2023 is the quarter when the open-source AI ecosystem becomes genuinely competitive with proprietary systems for a wide range of practical tasks. Llama 2, Code Llama, Mistral 7B, and Falcon 180B collectively establish that open-weight LLMs are not merely academic alternatives but production-deployable systems. Their commercial advantages are clear: no API costs, no data-sharing obligations, no single-vendor dependency. Enterprises and teams invest in understanding and deploying these open models in Q3 2023. They are building a capability that compounds over the next several years as the open ecosystem accelerates. Additionally, the organisations that wait for a perfect proprietary API solution are conceding a significant head start.
7. Summary
The third quarter of 2023 is the open-source AI summer: a concentrated three-month burst of open-weight model releases. That burst restructures the competitive landscape of artificial intelligence more profoundly than any comparable period since the GPT-3 API launch in 2020. Llama 2 ships on July 18 and immediately becomes the reference point against which all open-weight alternatives are measured. Code Llama arrives on August 24 to challenge GitHub Copilot on code generation. Mistral 7B lands on September 27, demonstrating two things through architectural elegance. Small, well-designed models beat larger competitors, and European AI research is a genuine frontier force. By the end of September, an engineer can run a GPT-3.5-class language model on a laptop. No network connection, cloud account, or per-token billing relationship is required.
On the proprietary frontier, the race continues at equal pace. Anthropic’s Claude 2 resets expectations for context window length at 100,000 tokens, enabling applications in legal review and document analysis that were previously impractical. OpenAI’s ChatGPT Enterprise resolves the data governance question that had prevented regulated enterprises from adopting ChatGPT. In addition, the August renaming of Code Interpreter to Advanced Data Analysis signals OpenAI’s ambition to own the analyst workflow, not just the developer workflow. Microsoft announces a November 1 general availability date and $30 per user per month pricing for Microsoft 365 Copilot. That establishes the first clear market rate for AI in enterprise productivity software.
7.1 Open questions and what to watch in Q4 2023
The open questions that remain at the end of Q3 2023 are significant. Furthermore, will open-weight models close the remaining benchmark gap with GPT-4 in 2024, or does frontier training require proprietary resources that no open effort can replicate? Moreover, how will the EU AI Act’s trilogue resolve the contentious debates around general-purpose AI model obligations? Will foundation model providers face direct compliance duties, or will liability sit entirely with the downstream application deployer? Additionally, can the GPU supply chain catch up with demand before the inference efficiency improvements make the supply constraint irrelevant for most applications?
Practitioners entering Q4 2023 should watch three developments with particular attention. First, consider the Mistral AI roadmap. This is a startup that ships a model beating Llama 2 13B in its first public release, under a genuinely permissive license. It is almost certain to produce follow-on models that further close the gap with proprietary systems. Second, Microsoft 365 Copilot’s actual enterprise adoption curve: the announced price and feature set are compelling. However, enterprise AI rollout depends heavily on change management, governance approvals, and IT security reviews that take months. Third, watch the vLLM and inference optimisation ecosystem. The 24x throughput improvement demonstrated by PagedAttention in production is a preview. Inference costs can fall dramatically as the engineering community focuses on the problem.
7.2 The question the quarter resolves
Q3 2023 resolves a fundamental question that has been open since the ChatGPT launch in November 2022. Is frontier AI the exclusive domain of a small number of extremely well-capitalised US companies? The answer, by the end of September 2023, is clearly no. However, the open-source AI summer proves that open-weight models are viable, competitive, and commercially deployable. Therefore, the next competitive battleground is not model capability alone but the combination of capability, governance, cost, and trust that enterprises require at scale.
8. Sources
- Meta — “Meta and Microsoft Introduce the Next Generation of Llama” – Official Meta announcement of Llama 2. July 18, 2023.
- Touvron et al. — “Llama 2: Open Foundation and Fine-Tuned Chat Models” – arXiv preprint. July 18, 2023. Technical details of Llama 2 architecture and training.
- Meta AI — “Introducing Code Llama, a state-of-the-art large language model for coding” – Official Meta AI blog post. August 24, 2023.
- Rozière et al. — “Code Llama: Open Foundation Models for Code” – arXiv preprint. August 24, 2023.
- Mistral AI — “Mistral 7B” – Official Mistral AI announcement post. September 27, 2023.
- Jiang et al. — “Mistral 7B” – arXiv preprint, October 2023. Technical paper describing GQA, sliding window attention, and benchmark comparisons.
TechCrunch — “Mistral AI makes its first large language model free for everyone”
Devin Coldewey. September 27, 2023.
Anthropic — “Claude 2”
Official Anthropic announcement. July 11, 2023. Claude 2 capabilities and context window.
OpenAI — “Introducing ChatGPT Enterprise”
Official OpenAI blog post. August 28, 2023.
ggerganov/llama.cpp on GitHub
MIT license. C/C++ inference engine for Llama-family models. GGUF format introduced Q3 2023.
langchain-ai/langchain on GitHub
MIT license. LangChain Expression Language (LCEL) introduced Q3 2023.
vLLM — High-throughput and memory-efficient inference engine for LLMs
UC Berkeley Sky Computing Lab. PagedAttention algorithm providing up to 24x throughput improvement.
Microsoft — “Announcing Copilot in Microsoft 365”
September 21, 2023. Pricing and GA date announcement for Microsoft 365 Copilot.
Amazon Web Services — “Amazon Bedrock”
Product page. Managed foundation model service covering Anthropic Claude, Meta Llama 2, and Stability AI models.
Hugging Face — tiiuae/falcon-180B model card
Technology Innovation Institute. Falcon 180B released September 6, 2023. 3.5T token RefinedWeb training.
Simon Willison — “Large language models are having their Stable Diffusion moment”
simonwillison.net. March 11, 2023.
simonw/llm on GitHub
Simon Willison’s command-line LLM tool. Apache 2.0 license.
Andrej Karpathy — YouTube channel
“Neural Networks: Zero to Hero” series including “Let’s build GPT from scratch.”
Sebastian Raschka — Ahead of AI newsletter
Technical deep dives on LLM architectures, LoRA, QLoRA, and open-weight model comparisons throughout Q3 2023.
European Parliament — “MEPs ready to negotiate first-ever rules for safe and transparent AI”
June 14, 2023. Parliament adopts negotiating position for EU AI Act.
Stability AI — “Stable Diffusion XL 1.0 is Here”
July 26, 2023. SDXL 1.0 open-weight image generation model release.
Ars Technica — “Meta launches LLaMA-2, a source-available AI model that allows commercial applications”
Benj Edwards. July 18, 2023.
Open Source Initiative — “Meta’s LLaMa license is not Open Source”
Stefano Maffulli. July 20, 2023. OSI critique of Llama 2 community license.
Apple Newsroom — “Apple unveils iPhone 15 and iPhone 15 Plus”
September 12, 2023. iPhone 15 announcement with A17 Pro chip and USB-C transition.
xAI — Official Company Website
xAI incorporated July 2023 by Elon Musk as a frontier AI research company.
Free ebook
Free AI Video, Generated Locally
Run Wan 2.1 in ComfyUI on your own GPU — the scripts I use, measured times, sample clips. No cloud, no API keys.
No spam. Unsubscribe at any time.
Leave a Reply