Q4 2023 IT Review – GPT-4 Turbo, Gemini, EU AI Act, and the OpenAI Crisis

Part of IT Trends & Reviews — what actually shipped, quarter by quarter.

Introduction: The Quarter AI Grew Up

Q4 2023 is the quarter when artificial intelligence stopped being a research curiosity. Instead, it became a product category, a regulatory challenge, and a governance crisis all at once. In the span of twelve weeks, OpenAI launched GPT-4 Turbo with a 128 000-token context window. Meanwhile, Google announced Gemini Ultra in a landmark multimodal demonstration. The EU reached a political agreement on the world’s first comprehensive AI Act. Mistral released Mixtral 8x7B into the open under an Apache 2.0 licence. Moreover, Anthropic shipped Claude 2.1 with halved hallucination rates. Finally, Sam Altman was fired and reinstated as OpenAI CEO within five tumultuous days.

This is not a single thread of progress. It is several simultaneous revolutions colliding. One is the race to extend context windows. Another is the push toward open weights. A third is the scramble to write law fast enough to catch up with capability. Finally, there is the realisation that the organisations building frontier AI face profound internal tensions. Those tensions run between commercial ambition and safety-first culture.

At the same time, the quarter produced a quieter but equally important signal: the physical world started catching up. For example, industrial IoT platforms began adopting edge AI for predictive maintenance at scale. Similarly, automated robotic systems moved into sectors as unglamorous as truck refuelling. Meanwhile, ABB opened an innovation lab in China to bring robotics research closer to its largest manufacturing market. In short, the AI revolution is no longer contained in the cloud.

Overall, this review covers the principal developments across all three dimensions: the model and tooling race, the regulatory inflection point, and the industrial edge. Moreover, it draws on publicly available announcements, GitHub release data, and contemporaneous reporting.

GitHub & Open Source Deep Dive

Q4 2023 is the quarter that confirmed open-source AI was not merely catching up with proprietary models — in some specialised dimensions it was pulling ahead. Notably, the most dramatic proof arrived on 11 December. Mistral AI released Mixtral 8x7B via a magnet link posted to X. There was no prior warning, no press release, and no blog post. The model used a sparse mixture-of-experts (MoE) architecture. In practice, each layer held eight expert feed-forward networks, and each token was routed to just two. That yielded the effective parameter count of a 12.9 B dense model. Meanwhile, inference cost stayed comparable.

Performance benchmarks were startling. Mixtral 8x7B matched or exceeded LLaMA 2 70B on most standard evaluations while consuming roughly a third of the compute. Notably, on code-generation benchmarks it outperformed GPT-3.5. Mistral made the weights available under Apache 2.0 — a decision that immediately triggered a wave of fine-tunes. It also triggered quantised GGUF builds for consumer hardware and integration PRs across every major inference framework.

LangChain and the Orchestration Ecosystem

Meanwhile, LangChain v0.1 crossed 70 000 GitHub stars during the quarter, cementing its position as the default orchestration framework for production LLM applications. In addition, the project released LangServe, a FastAPI-based serving layer that converts any LangChain runnable into a REST API with automatic input/output schema generation. As a result, this dramatically lowered the barrier between prototype and deployment.

Furthermore, LangChain Expression Language (LCEL) reached stability in Q4, offering a composable pipe-based syntax for constructing chains and enabling streaming by default. The community celebrated. However, critics noted that the abstraction occasionally obscured what was happening underneath. The maintainers acknowledged that complaint by shipping improved tracing and debugging tooling alongside.

AutoGen: Multi-Agent Coordination Arrives

Meanwhile, Microsoft Research released AutoGen to the public in Q4 2023, a framework for orchestrating multiple cooperating AI agents to solve complex tasks through conversation. AutoGen enables patterns where a human-in-the-loop interacts with a group of specialised agents. For example, that group can include a planner, a coder, a verifier, and a critic. Together they negotiate toward a solution. For instance, early demos showed AutoGen solving competitive programming problems by iterating between a code-writing agent and a test-running agent.

The release crystallised “agentic AI” as a genuine engineering discipline. Prior discussion of AI agents had been largely conceptual; AutoGen provided an opinionated framework with concrete primitives: UserProxyAgent, AssistantAgent, group chats, and configurable execution environments. Overall, by end of quarter the repo had accumulated over 20 000 stars.

Key Repository Milestones

Repository	Q4 2023 Milestone	Significance
mistralai/mistral-src	Mixtral 8x7B weights released (Apache 2.0)	Best open-weight model vs. compute ratio to date
microsoft/autogen	Public release, 20 000+ stars	First mainstream multi-agent framework
langchain-ai/langchain	v0.1 stable, LangServe, LCEL	Orchestration layer matures toward production
ggerganov/llama.cpp	Mixtral GGUF support within 48 hours	Consumer hardware LLM inference normalised
openai/openai-python	v1.0 rewrite with async support	Aligned with new Assistants API surface

Big Tech Moves

OpenAI DevDay: GPT-4 Turbo and the Assistants API

On 6 November 2023, OpenAI held its first developer conference, DevDay, in San Francisco. Above all, the headline announcement was GPT-4 Turbo. The model offered a 128 000-token context window, equivalent to roughly 300 pages of text. It also carried a training knowledge cutoff of April 2023. Moreover, pricing fell significantly. Input tokens dropped from $30 to $10 per million, and output from $60 to $30. The cost reduction was not incremental; it was transformative for use-cases that had been economically marginal with GPT-4.

In addition, OpenAI launched the Assistants API. This new surface manages conversation state, retrieval from uploaded files, and code execution via a sandboxed interpreter. It also handles function calling in a single stateful object. By comparison, developers previously had to manage threads, embeddings, and tool dispatch themselves. Thus, the Assistants API collapsed four separate integrations into one. Custom GPTs added a no-code layer for building specialised chat interfaces. A GPT Store was planned for early 2024. Combined, these moves showed OpenAI visibly attempting to become a platform, not just an API.

Overall, the subtext of DevDay was capability democratisation. GPT-4-level intelligence now costs ten times less and can hold an entire novel in context. Therefore, the competitive moat for applications built on raw model access effectively disappears. Instead, the value shifts to data, distribution, and product design.

The OpenAI Leadership Crisis

Eleven days after DevDay, on 17 November, OpenAI’s board fired CEO Sam Altman. The board offered no public explanation beyond a statement that he had not been “consistently candid” with the board. In addition, President Greg Brockman resigned the same day. Notably, the company’s chief scientist Ilya Sutskever, a board member, signed the letter but later expressed regret.

What followed was an extraordinarily public corporate implosion. Virtually the entire OpenAI staff — reportedly over 700 employees — signed a letter threatening to resign unless Altman was reinstated and the board resigned. Meanwhile, Microsoft CEO Satya Nadella announced that Altman and Brockman would join Microsoft to lead a new AI research group if reinstatement did not occur. The spectacle of the world’s most commercially significant AI company potentially dissolving over a weekend concentrated the mind of the entire industry.

Free ebook

Free AI Video, Generated Locally

Working scripts and measured benchmarks. Free.

Your e-mail address

Leave this field empty

Get the ebook

No spam. Unsubscribe at any time.

Finally, by 21 November, Altman was reinstated. In addition, a reconstituted board took over, with several original members, including Helen Toner and Tasha McCauley, departing. The episode exposed a structural tension that had always been implicit in OpenAI’s unusual corporate structure. In practice, a non-profit board held ultimate control over a capped-profit subsidiary. That subsidiary’s commercial success had already made some employees paper-millionaires. The question of whose interests the board represented — humanity’s, or the company’s — was answered empirically by the market within five days.

Google Gemini

Meanwhile, on 6 December, Google DeepMind unveiled Gemini, its most capable model family to date, in three sizes: Ultra, Pro, and Nano. The announcement was accompanied by a benchmark-rich research paper. In addition, a striking video demonstration showed Gemini responding to handwritten text, hand-drawn diagrams, and spoken queries in real time. Gemini Ultra reportedly outperformed GPT-4 on 30 of 32 standard academic benchmarks. It was also the first model to surpass human expert performance on MMLU (Massive Multitask Language Understanding).

Meanwhile, Gemini Pro was made available via Google AI Studio and the Vertex AI platform immediately after the announcement. Gemini Nano was deployed directly on-device to the Pixel 8 Pro. For example, it enabled features like Summarise in Recorder and Smart Reply in Gboard. Those features ran entirely locally without network round-trips. This on-device deployment was a deliberate signal. AI inference could be private, low-latency, and energy-efficient. That holds when the model architecture is designed for it from the ground up.

However, a week after launch, questions emerged about the promotional demo video’s editing. The video had been compiled from individual interactions rather than a live session. In addition, some prompts shown in the video were more explicit than the ones actually used during filming. Google acknowledged the editing; the response in the developer community ranged from mild disappointment to pointed criticism. Nevertheless, it did not materially change Gemini’s technical standing. However, it crystallised a pattern of AI companies overstating capability in launch materials.

Microsoft: Copilot Everywhere

Meanwhile, Microsoft shipped GitHub Copilot Chat to all Copilot subscribers in October. It extended beyond code completion to conversational code assistance, test generation, and commit message suggestion. Those capabilities integrated directly into Visual Studio and VS Code. In addition, Copilot for Microsoft 365 brought GPT-4 to Word, Excel, Outlook, and Teams. It became available to all enterprise customers at $30 per user per month. Previously it had been restricted to a limited early access programme.

Overall, the breadth of Microsoft’s Copilot rollout in Q4 was remarkable. By December 2023, there was a Copilot variant for Azure, helping write infrastructure code and queries. Another served Power Platform, generating business logic from natural language. A third served Security, correlating threat intelligence. A fourth served Dynamics 365, summarising customer interactions. In short, Microsoft was effectively rebundling its entire product portfolio around a GPT-4 runtime.

NVIDIA: The Infrastructure Backbone

Meanwhile, NVIDIA reported Q3 fiscal 2024 results in November with data centre revenue of $14.5 billion. That was a 279% year-over-year increase. It confirmed that the GPU supply chain was the primary constraint on AI development globally. At the same time, the H100 remained the dominant training accelerator. In addition, NVIDIA announced the H200, incorporating HBM3e memory with 141 GB capacity and 4.8 TB/s bandwidth. Delivery was targeted for Q1 2024.

Moreover, NVIDIA launched TensorRT-LLM, an open-source library for optimising LLM inference on NVIDIA hardware, with pre-built engines for LLaMA 2, Mistral, Falcon, and other popular architectures. TensorRT-LLM complemented the vLLM project from UC Berkeley. That project had emerged earlier in 2023 as the preferred high-throughput serving framework for open-weight models. It saw substantial adoption growth in Q4 following the Mixtral release.

AI Impact & Tooling Explosion

Anthropic Claude 2.1

Anthropic released Claude 2.1 in November with a 200 000-token context window. That was the largest available in any commercial model at the time. The release also added a system prompt tool-use capability. It allowed developers to grant Claude access to specific functions. Anthropic also published updated evaluation data showing a 50% reduction in hallucination rates compared to Claude 2.0 on their internal factual accuracy benchmarks.

Claude 2.1’s long context was not merely a headline figure. Anthropic demonstrated it could accurately answer questions from documents positioned near the middle of a 200K-token context. In that regime, GPT-4 (at 128K) had well-documented “lost in the middle” degradation. For use cases involving long legal documents, codebases, or extended research reports, Claude 2.1 represented a genuine capability advantage.

The Context Window Wars

Q4 2023 marks the quarter when context window size became a primary competitive dimension. The arc across the year had moved from GPT-4’s 8K (launch) to 32K (extended), to Claude 2.0’s 100K, to GPT-4 Turbo’s 128K, to Claude 2.1’s 200K. Each extension unlocked new application categories: entire codebases could be loaded, entire books could be analysed, long multi-party conversations could be maintained without summarisation hacks.

The engineering challenge behind longer contexts is quadratic: standard transformer attention scales as O(n²) with sequence length. The techniques deployed in Q4 included rotary position embeddings with extended base frequencies (RoPE scaling), sliding window attention (Mistral), and various forms of sparse attention. None was universally dominant; the field was exploring the tradeoff space between memory cost, latency, and quality degradation at extreme lengths.

Retrieval-Augmented Generation Matures

RAG — the pattern of retrieving relevant documents at inference time and including them in the prompt — matured considerably in Q4. LlamaIndex released a substantial update introducing agents that could reason over graph-structured knowledge, not just flat vector stores. The concept of “agentic RAG” emerged. Instead of a single retrieve-then-generate step, an agent could decide what to retrieve. It could then evaluate the retrieved content and issue follow-up retrieval queries if the first results were insufficient.

Several startups built dedicated vector database products in Q4, including continued growth at Pinecone, Weaviate, Qdrant, and Chroma. Notably, the Assistants API’s built-in retrieval tool handled chunking, embedding, and search automatically. It therefore put pressure on the standalone vector database narrative. If OpenAI abstracts away the complexity, why build bespoke infrastructure? The answer for enterprise users remained data residency, latency control, and avoiding vendor lock-in.

EU AI Act: The Regulatory Inflection Point

On 8 December, European Parliament and Council negotiators reached a political agreement on the EU AI Act after marathon trilogue negotiations. The Act introduces a risk-tiered framework. Prohibited uses include real-time biometric surveillance in public spaces and social scoring. High-risk systems, such as medical devices, critical infrastructure, and hiring algorithms, require conformity assessments and post-market monitoring. Finally, general-purpose AI (GPAI) models above a certain compute threshold face transparency and safety evaluation requirements.

The GPAI provisions were the most contested. Frontier AI developers argued that compute-based thresholds were a poor proxy for capability or risk, that evaluation requirements were not yet scientifically well-defined. In addition, they argued that mandatory disclosure of training data would expose trade secrets. European civil society organisations argued the provisions were too weak and that self-assessment by developers was insufficient for systems capable of generating disinformation at scale.

The agreement was politically significant beyond its immediate legal effect. It established that the EU was prepared to regulate frontier AI through binding law, not voluntary codes of conduct. Other jurisdictions — the UK, Canada, Brazil, and several Asian economies — were watching closely. The US approach via executive order, announced October 30, was process-oriented rather than substantive, directing agencies to develop guidelines and standards rather than imposing requirements directly. The regulatory divergence between jurisdictions became a strategic consideration for AI companies planning global deployments.

Key Voices & Releases

Trend Synthesis

1. The Context Window as Platform Primitive

Context windows expanded rapidly in Q4 2023, from tens of thousands to hundreds of thousands of tokens. That is not just a performance improvement. It changes the architecture of AI applications. When the context window is large enough to contain an entire codebase, a full legal case, or a product’s complete documentation, the retrieval step becomes less critical. The application pattern simplifies: load everything, ask the question, get the answer. This pushes value away from retrieval infrastructure and toward the quality of the underlying model. It also enables entirely new use cases: real-time collaboration on long documents, code refactoring with full project awareness, and multi-document synthesis without lossy chunking.

2. Open Weights Change the Competitive Map

Mixtral 8x7B’s surprise December release demonstrated that the open-weight ecosystem was no longer playing catch-up on an indefinite lag. Apache 2.0 licensing means these weights can be used commercially without restriction. Some enterprises are uncomfortable with data passing through third-party APIs. Others work in regulated industries or sensitive domains. For them, a locally-deployable model of GPT-3.5 quality changes the options analysis fundamentally. Mistral’s strategy — open weights to build community and reputation, proprietary fine-tuned variants and managed API for revenue — echoes the strategy of successful open-core software businesses.

3. AI Governance Is Now Real

The OpenAI board crisis and the EU AI Act political agreement occurred within three weeks of each other. Jointly, they establish that AI governance is no longer an abstract future concern. It is the present. The crisis showed that even the most commercially successful AI organisation has unresolved questions. Those questions concern accountability structures, board composition, and the relationship between safety culture and commercial execution. The EU Act showed that democratic legislatures will write rules, even imperfect ones, when capability advances faster than voluntary norms. The combination of corporate governance failure and regulatory action in the same quarter signals a permanent shift. AI development is now contextualised as an activity with legal, fiduciary, and societal obligations. Those obligations are not just technical and commercial ones.

4. IIoT and Edge AI Are the Industrial AI Layer

Industrial adoption is a lagging but real signal of where AI deployment is actually happening in manufacturing and logistics. IIoT platforms like OPTIFY, automated physical systems like Autofuel, and robotics innovation centres like ABB Shanghai represent one deployment wave. That wave is more durable than software-only AI products. It is embedded in physical capital with 10-to-20-year lifecycles. A bearing condition monitoring sensor installed in a grain elevator in 2023 will still be generating data in 2035. The companies that build the platforms to aggregate and model that data are building defensible positions that foundation model providers cannot easily replicate.

5. The Orchestration Layer Becomes Infrastructure

LangChain, AutoGen, LlamaIndex, and similar frameworks moved from developer toys to production infrastructure in Q4. The patterns they encode — retrieval-augmented generation, tool-calling, multi-agent coordination, human-in-the-loop approval — are emerging as standard components of AI application architecture. Whether these specific frameworks persist or are replaced by later iterations, the patterns themselves are becoming the vocabulary of AI engineering. Teams that understand these patterns deeply will be better positioned than teams that merely know how to call an API.

Trend	Q4 2023 Proof Point	Direction
Context window expansion	GPT-4 Turbo 128K; Claude 2.1 200K	Accelerating
Open-weight quality	Mixtral 8x7B matches LLaMA 2 70B	Accelerating
AI regulation	EU AI Act political agreement	Emerging
Corporate AI governance	OpenAI board crisis	Critical
IIoT / edge AI	OPTIFY, Autofuel, ABB Shanghai	Maturing
Multi-agent frameworks	AutoGen, agentic RAG, LangServe	Early growth

Summary

Q4 2023 delivers more structural change in a single quarter than most technology domains see in a decade. The AI stack is simultaneously getting more capable, more open, more industrially deployed, and more regulated. Capability comes from GPT-4 Turbo, Gemini, and Claude 2.1. Openness comes from Mixtral 8x7B under Apache 2.0. Industrial deployment comes from OPTIFY, Autofuel, and ABB. Regulation comes from the EU AI Act and the US Executive Order. The OpenAI governance crisis adds a fifth dimension: organisations building transformative technology cannot defer questions of accountability and oversight indefinitely.

For practitioners, the actionable takeaways from Q4 2023 are concrete. Context window expansion removes a major design constraint; applications that were architecturally complex due to context limits can be simplified. Open-weight models at GPT-3.5 quality change the build-versus-API calculus for privacy-sensitive or latency-critical applications. The orchestration patterns — RAG, tool-calling, multi-agent — are stable enough to build production systems on. And the IIoT signal carries a reminder. AI value creation in 2023 and beyond is not confined to generative models. The industrial deployment layer is less glamorous and less discussed. Yet it is where AI becomes physically permanent.

Q4 2023 is a hinge quarter. The decisions made now — about model access, governance structures, regulatory compliance, and industrial deployment strategy — will shape the AI landscape for years. Notably, the pace of change is not slowing. The question is whether institutions, organisations, and regulations can develop at a compatible speed.

Sources

1. OpenAI. OpenAI DevDay: New models and developer products announced. November 6, 2023. https://openai.com/blog/new-models-and-developer-products-announced-at-devday
2. OpenAI. OpenAI announces leadership transition. November 17, 2023. https://openai.com/blog/openai-announces-leadership-transition
3. OpenAI. Sam Altman returns as CEO, OpenAI has a new initial board. November 22, 2023. https://openai.com/blog/sam-altman-returns-as-ceo-openai-has-a-new-initial-board
4. Google DeepMind. Introducing Gemini: our largest and most capable AI model. December 6, 2023. https://deepmind.google/technologies/gemini/
5. Mistral AI. Mixtral of experts. December 11, 2023. https://mistral.ai/news/mixtral-of-experts/
6. Anthropic. Claude 2.1. November 21, 2023. https://www.anthropic.com/news/claude-2-1
7. European Parliament. EU AI Act: deal on comprehensive rules for trustworthy AI. December 9, 2023. https://www.europarl.europa.eu/news/en/press-room/20231206IPR15699/eu-ai-act-deal-on-comprehensive-rules-for-trustworthy-ai
8. Microsoft Research. AutoGen: Enabling next-generation large language model applications. 2023. https://github.com/microsoft/autogen
9. LangChain AI. LangChain v0.1.0. 2023. https://github.com/langchain-ai/langchain/releases
10. NVIDIA Investor Relations. NVIDIA announces financial results for third quarter fiscal 2024. November 21, 2023. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-third-quarter-fiscal-2024
11. Dodge Industrial. OPTIFY Condition Monitoring Platform. 2023. https://www.dodgeindustrial.com/en-US/products/digital-solutions/optify
12. ABB. ABB Robotics opens Open Innovation Lab in Shanghai. October 2023. https://new.abb.com/news/detail/111016/abb-robotics-unveils-open-innovation-lab
13. Microsoft. GitHub Copilot Chat is now generally available. October 2023. https://github.blog/2023-12-29-github-copilot-chat-now-generally-available-for-organizations-and-individuals/
14. White House. Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. October 30, 2023. https://web.archive.org/web/20231115013140/https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/

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.

Your e-mail address

Leave this field empty

Get the ebook

No spam. Unsubscribe at any time.