Q3 2024 IT Quarterly Review – Llama 3.1, OpenAI o1, CrowdStrike, EU AI Act

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

1. Introduction – The Quarter of Two Philosophies

July 2024 opens with a paradox that the industry spends the rest of the quarter trying to resolve. In fact, within five days, Meta releases the largest open-weights language model in history. It has 405 billion parameters and is freely downloadable, tradeable, and deployable on any hardware you can afford. OpenAI meanwhile ships GPT-4o mini, a compact model that undercuts every competitor on cost-per-token while remaining firmly closed-source. The message from two of the world’s most influential AI labs is irreconcilable on the surface. In short, one bet is that openness wins, the other that quality and control are worth guarding.

Underneath this philosophical contest, Q3 2024 is the quarter when artificial intelligence stops being the property of researchers and becomes infrastructure. In practice, the shift is visible in the data. For example, GPT-4o mini replaces GPT-3.5 Turbo across OpenAI’s platform at a fraction of the cost. Capable AI becomes available to student projects, two-person startups, and hobby APIs at $0.15 per million input tokens. Llama 3.1 405B achieves performance benchmarks that match GPT-4-class models, for the first time in the open-weights world. Any organization with sufficient compute can therefore train a frontier model on its own proprietary data, without ever transmitting that data to a third-party API. As a result, the tectonic plates of enterprise AI procurement shift visibly over the course of ninety days.

1.1 CrowdStrike and the AI Act

Then July 19 happens. A single malformed content configuration file from a CrowdStrike sensor update pushes a kernel-mode driver fault to 8.5 million Windows devices simultaneously. Airlines cancel thousands of flights. Hospitals revert to paper records. Banks go offline. Notably, the outage is not caused by a cyberattack, nor by a deliberate act. It is the logical endpoint of a trend years in the making. In practice, security software with kernel-level privileges is silently auto-updated and deployed at scale, with minimal rollback capability. Q3 2024 is the quarter that demonstrates, at civilizational cost, that software resilience is not a DevOps footnote — it is geopolitical infrastructure.

Against this backdrop, the European Union’s Artificial Intelligence Act enters into force on August 1. In fact, it passed the European Parliament in March and appeared in the Official Journal of the European Union on July 12. The regulation’s implementation timeline is staged. Prohibitions on unacceptable-risk AI take effect in twelve months, and obligations for high-risk systems in twenty-four, which gives industry a runway but also a deadline. As a result, every software team working on AI-augmented products serving European users is now building to a compliance calendar, whether they know it or not.

1.2 o1 and the reasoning turn

OpenAI rounds out the quarter on September 12 with o1 preview. The model does not just generate text but reasons through problems, iterating on internal chains of thought before producing output. In practice, the model trades raw token speed for accuracy on complex reasoning, mathematics, and coding tasks. Consequently, its release forces a rethink of how AI products are priced and positioned. September 24 brings Advanced Voice Mode to ChatGPT Plus and Team subscribers in the United States. It completes a roadmap first previewed at the GPT-4o launch in May. Meta closes the quarter on September 25 with Llama 3.2. Notably, it adds multimodal vision capability and on-device models of 1B and 3B parameters optimized for mobile and edge deployment.

1.3 An industry entering its second phase

The overall picture is of an industry accelerating into its second phase. It moves from proof-of-concept assistants to embedded runtime components in the products and services that run civilization. The developer tooling around AI — local inference engines, RAG frameworks, evaluation harnesses, prompt management libraries — reaches a new level of maturity. For example, GitHub Copilot adds support for multiple underlying models. Similarly, Cursor, an AI-native fork of VS Code, begins attracting serious professional adoption. The question in Q3 2024 is no longer whether AI can do useful work. Rather, it is who controls the infrastructure that makes it run.

2. GitHub Deep Dive – Open-Source AI Infrastructure Explodes

The Llama 3.1 release triggers the most concentrated burst of open-source AI tooling activity since the original Llama leak of March 2023. Repositories that enable local inference, model quantization, retrieval-augmented generation, and multi-model routing all see dramatic star growth over the quarter. For developers who want to understand the gravitational center of applied AI in Q3 2024, watching GitHub tells the story faster than any press release.

2.1 meta-llama/llama-models – The Open-Weights Watershed

The meta-llama/llama-models repository is the canonical distribution point for Meta’s Llama 3.1 weights. Moreover, the July 23 release ships three model sizes — 8B, 70B, and 405B parameters — alongside corresponding instruction-tuned and chat variants. Furthermore, for each size, Meta provides model cards, system prompt specifications, and the Llama 3.1 Community License Agreement that governs use.

The technical advances in Llama 3.1 over its predecessor are significant. The context window expands from 8,192 tokens to 128,000 tokens, matching or exceeding GPT-4 Turbo’s context offering. As a result, retrieval architectures can now ingest and reason over entire codebases or legal documents in a single pass. The training dataset reaches 15 trillion tokens, and Llama 3.1 405B requires 440,000 GPU hours of training compute. Moreover, Meta makes the figure public, and it illustrates the staggering cost moat around models of this size. The model’s performance on MMLU, HumanEval, MATH, and multilingual benchmarks positions it credibly alongside GPT-4 class models. That prompts a productive debate across academic preprint servers, Hacker News, and AI engineering communities about whether the era of closed-source model dominance is ending.

On September 25, the same repository receives the Llama 3.2 weights: 1B, 3B, 11B, and 90B parameter models. In practice, the 1B and 3B variants target on-device inference on modern smartphones and edge accelerators. The 11B and 90B multimodal models are the first open-weights models from Meta that accept image input alongside text. They train on 9 trillion tokens with a 128K context window.

2.2 ggerganov/llama.cpp – Making 405B Runnable on Consumer Hardware

The release of a 405B model creates an immediate practical problem: almost no one owns hardware capable of running it at full precision. Georgi Gerganov’s llama.cpp project is a C++ implementation of transformer inference optimized for CPU and GGUF quantization formats. Consequently, it becomes essential infrastructure within hours of the Llama 3.1 release. Community contributors publish 4-bit quantized GGUF versions of all three Llama 3.1 sizes on Hugging Face within days of the weights appearing. A 4-bit quantized Llama 3.1 405B requires approximately 220 GB of RAM. That said, it is still well beyond a typical developer laptop, but achievable on a single multi-GPU server or a Mac Studio with maximum RAM configuration.

The GGUF file format stores both tensor data and model metadata in a single portable file. As a result, that enables rapid sharing and deployment across operating systems without Python runtime dependencies. In Q3 2024, llama.cpp adds support for speculative decoding. In that technique a smaller draft model proposes tokens rapidly and a larger verifier model accepts or rejects them. Consequently, it brings inference throughput on local hardware measurably closer to API response latencies. Meanwhile, the project star count crosses 65,000 during the quarter and the community pull-request rate accelerates substantially.

2.3 ollama/ollama – The Docker of Local AI

Ollama packages the llama.cpp backend with a clean CLI and REST API, making local model deployment feel like docker pull followed by docker run. The ollama run llama3.1:70b command pulls the model, manages hardware detection, and starts an OpenAI-compatible API endpoint. As a result, the workflow dramatically lowers the barrier to self-hosted AI for engineering teams. Notably, Q3 2024 brings Llama 3.1 support to Ollama within hours of the Meta release. In addition, the project’s model library expands to include dozens of community-contributed variants. Ollama gains traction not only among individual developers but among enterprises. They need a managed local inference solution compliant with data-sovereignty requirements, particularly in regulated industries where data cannot leave on-premises environments.

2.4 microsoft/graphrag – Structured Knowledge Meets Retrieval

Microsoft Research releases graphrag as an open-source project, implementing a retrieval-augmented generation approach that builds a knowledge graph from a document corpus before answering queries. Unlike naive vector-search RAG, GraphRAG extracts entities, relationships, and community summaries from documents. In practice, the system can then answer complex multi-hop questions that flat vector search handles poorly. One example is “what are the common themes across all customer complaint reports from the last fiscal year?”. Moreover, the project is built with asyncio and LangChain-compatible abstractions, making it accessible to AI engineering teams already familiar with the Python ML ecosystem. Its release generates significant discussion about a tradeoff. Graph construction carries a setup cost, set against the quality improvement on complex analytical queries that characterize most enterprise use cases.

2.5 BerriAI/litellm – One API to Rule Them All

The number of production-quality LLM providers expands to include OpenAI, Anthropic, Google Gemini, Cohere, Mistral, and now locally-hosted Llama variants. Engineering teams therefore face the operational problem of managing multiple SDKs, authentication schemes, rate-limit handling strategies, and token counting implementations. LiteLLM provides a single Python interface that routes to any of these providers with an OpenAI-compatible function signature. In addition, it adds cost tracking, retry logic, and load balancing across multiple API keys. Furthermore, a local proxy server drops into existing OpenAI SDK integrations. In Q3 2024, LiteLLM adds native support for Llama 3.1 via Ollama, Together AI, and Groq. Its adoption in production enterprise deployments grows substantially, as teams begin building provider-agnostic AI stacks that can switch models without application-level code changes.

2.6 simonw/llm – Command-Line AI for Power Users

Simon Willison’s llm CLI tool grows substantially in Q3 2024. In addition, it adds plugin support for Llama 3.1 via Ollama, Groq’s ultra-fast inference API, and Anthropic Claude 3.5 Sonnet. The tool provides a uniform command-line interface for prompting, piping text through models, storing conversation history in SQLite, and running embedding operations. It is a versatile Swiss Army knife for developers who want AI capabilities in shell scripts and terminal workflows. Notably, the plugin architecture allows community contributors to add providers rapidly. In addition, Willison blogs and explains the project’s documentation extensively himself. That makes it unusually approachable for developers new to the LLM ecosystem who want to explore model capabilities without building full application scaffolding.

3. Big Tech & Industry Breakthroughs

Q3 2024 is the most event-dense quarter for enterprise technology since the GPT-4 launch of March 2023. The three-month window includes a historic open-source model release, the largest global IT outage in history, and three significant OpenAI product launches. It also covers the entry into force of a landmark AI regulation, a major Apple platform update, and NVIDIA’s ongoing GPU supply ramp. In short, each would be the headline story in a less crowded quarter.

Date	Company / Entity	Event	Significance
Jul 18	OpenAI	GPT-4o mini launch	Replaces GPT-3.5 Turbo; $0.15/1M input tokens
Jul 19	CrowdStrike / Microsoft	Global Windows outage	8.5M devices down; largest IT incident in history
Jul 23	Meta AI	Llama 3.1 release	First open-weights model matching GPT-4; 405B parameters
Aug 1	European Union	EU AI Act enters into force	World’s first comprehensive AI law; phased compliance timeline
Aug 20	OpenAI	GPT-4o fine-tuning for enterprise	Custom model training on proprietary data via API
Sep 12	OpenAI	o1 preview and o1-mini release	Reasoning model with internal chain-of-thought; PhD-level STEM
Sep 24	OpenAI	ChatGPT Advanced Voice Mode	Low-latency spoken conversation for Plus/Team users in US
Sep 25	Meta AI	Llama 3.2 multimodal release	1B/3B for on-device; 11B/90B with vision capability

3.1 OpenAI – GPT-4o Mini, Fine-Tuning, and the o1 Reasoning Model

OpenAI ships three major product updates across Q3 2024 that collectively reshape how developers and enterprises think about AI model selection. The first, GPT-4o mini on July 18, is a small multimodal model that replaces GPT-3.5 Turbo as the default low-cost option across OpenAI’s platform. The model costs $0.15 per million input tokens, roughly sixty times cheaper than GPT-4o at launch. As a result, it enables AI-powered features in applications where the economics of a premium model were previously prohibitive. GPT-4o mini supports the same 128K context window as the full GPT-4o, and it processes images. It also achieves higher benchmark scores than GPT-3.5 Turbo on standardized tests including MMLU and MGSM. That makes the deprecation of the older model straightforward for the vast majority of existing use cases.

The second update, on August 20, unlocks GPT-4o fine-tuning for enterprise API customers. In practice, organizations can submit training examples via the API, run supervised fine-tuning jobs on their proprietary datasets, and deploy custom model checkpoints at inference. This capability has long been available for GPT-3.5 Turbo and earlier models. It reaches GPT-4o class performance for the first time. Building domain-specialized models that match company style guides, terminology, and decision-making patterns becomes feasible without the cost and complexity of training from scratch. At the same time, the initial rollout targets safety-vetted enterprise customers, with broader availability planned for subsequent months.

3.2 OpenAI: the o1 reasoning model

The third update is the most consequential architecturally. On September 12, OpenAI releases o1 preview and o1-mini, models that represent a new paradigm in LLM product design. GPT-4o generates tokens in a single forward pass, which is fast, cheap, and suitable for most tasks. o1 instead spends time thinking before answering. In practice, it runs an internal reinforcement-learning-trained chain of thought that is not exposed to the user.

The result is a model that solves International Mathematical Olympiad problems at rates that match top human performers. Moreover, it scores at the 89th percentile on competitive programming platforms. In addition, it achieves PhD-level performance on physics, chemistry, and biology benchmarks. The cost and latency tradeoff is explicit: o1 is slower and more expensive than GPT-4o. In addition, the thinking tokens are priced separately. But for applications where correctness matters more than speed — financial modeling, legal analysis, scientific literature review — o1 establishes a new capability threshold.

3.3 Meta AI – Llama 3.1 and the 405B Bet on Open Weights

A strategic statement from CEO Mark Zuckerberg accompanies Meta’s Llama 3.1 release on July 23. Notably, it frames open-source AI as the correct long-run bet for both Meta and the technology industry. Zuckerberg makes the case in a widely shared blog post and a lengthy interview with Dwarkesh Patel the following day. In short, he argues that the value of AI is captured downstream. It lands in products, personalization, and integration with existing services, rather than in model weights themselves. Meta’s bet is that giving the weights away for free generates a larger ecosystem of developers, fine-tuners, and application builders. Consequently, they collectively improve the model family faster than any closed-source team working alone.

The 405B model is trained at an estimated cost comparable to the annual R&D budget of a mid-size software company. Nevertheless, it becomes publicly downloadable within hours of the announcement. In addition, Meta publishes a technical paper co-authored by dozens of researchers, detailing the pre-training data pipeline, the supervised fine-tuning approach, and the safety evaluation protocols. The model ships under the Llama 3.1 Community License, which permits commercial use subject to usage restrictions. Its restrictions are subsequently characterized by the Open Source Initiative as falling outside the official Open Source Definition. The debate generates productive friction about what “openness” means for foundation models with enormous downstream societal impact.

3.4 Meta: Llama 3.2 and on-device models

Llama 3.2, released on September 25 at Meta Connect, extends the family in two directions simultaneously. It reaches downward toward on-device deployment with the 1B and 3B parameter models optimized for ARM processors. It also reaches horizontally into multimodal capability with the 11B and 90B vision-language models. Notably, the vision models accept image input natively and are trained on 9 trillion tokens including captioned image data. This makes Llama 3.2 the first open-weights model family with both mobile-scale efficiency and frontier-adjacent multimodal capability. The combination matters enormously for use cases ranging from accessibility applications to industrial quality inspection systems operating at the network edge.

3.5 CrowdStrike – When Software Update Becomes Global Catastrophe

On July 19, 2024, at approximately 04:09 UTC, CrowdStrike pushes a content configuration update to its Falcon sensor software running on Windows hosts worldwide. The update contains a logic error in a Channel File. It instructs the sensor’s kernel-mode driver to parse an array of twenty-one pointers, but the file provides only twenty. The result is an out-of-bounds read, a kernel panic, and a non-bootable Blue Screen of Death on every affected device. Recovery requires booting each machine into safe mode and manually deleting the offending configuration file. In practice, the process takes thirty minutes per machine on average. It can take days in environments with encrypted drives or BitLocker recovery keys stored in inaccessible cloud portals. Large server fleets in data centers without automated remediation tooling are equally slow.

The scale of the incident is unprecedented: 8.5 million devices affected, according to Microsoft’s published analysis. For instance, airlines cancel or delay more than 5,000 flights on the day of the incident. Hospital systems in the United States, United Kingdom, and Australia divert patients and revert to paper-based procedures. Similarly, the London Stock Exchange experiences trading disruptions. Meanwhile, emergency services in multiple countries operate in degraded mode for hours. Insurance industry estimates put the economic loss in the range of $5 billion, making it the costliest software-related disruption in history.

3.6 CrowdStrike: what the incident teaches

The incident enters the technical literature as a case study in the inherent risks of kernel-level security software. Such software auto-updates with insufficient staged rollout, automated testing, and rollback capability. The error is not in executable code but in a data file. Such files are not typically subject to the same rigorous review cycle as compiled modules. That highlights a gap in enterprise software governance which applies far beyond CrowdStrike’s specific implementation. As a result, the aftermath drives renewed interest in immutable infrastructure and automated canary deployments. Hardware security module architectures that limit the blast radius of software update failures at any layer of the stack draw attention too.

3.7 European Union – The AI Act Enters into Force

It appears in the Official Journal of the European Union on July 12, 2024, and enters into force twenty days later, on August 1. The regulation establishes a tiered risk framework for AI systems. First, unacceptable-risk applications are prohibited entirely within twelve months of entry into force. They cover social scoring by governments, real-time biometric surveillance in public spaces, and subliminal manipulation. Second, high-risk applications cover critical infrastructure, education, employment, essential services, law enforcement, border control, and the administration of justice. They must meet conformity assessment, logging, transparency, and human oversight requirements within twenty-four months.

General-purpose AI models with “systemic risk” — defined initially as models trained on compute exceeding 10²⁵ FLOPs — face additional transparency and safety obligations. This threshold captures the current frontier of GPT-4 and Llama 3.1 scale training runs. In addition, providers must conduct red-team evaluations, maintain model documentation, and report incidents to national authorities. That said, the compliance timeline is long enough that most affected organizations are still in the risk assessment phase by the end of Q3 2024. However, the awareness that a legal framework now exists changes procurement conversations in financial services and healthcare sectors where AI adoption is accelerating fastest.

3.8 NVIDIA – Blackwell Architecture and the GPU Supercycle

NVIDIA’s Blackwell architecture, announced at the GTC conference in March 2024, moves toward production deployment during Q3. The GB200 NVL72 rack-scale system combines 36 Grace CPU chips, 72 Blackwell B200 GPUs, and NVLink interconnects capable of 1.8 TB/s chip-to-chip bandwidth. As such, it represents a generational step in AI training and inference density. However, supply constraints mean that initial GB200 shipments go to hyperscale cloud providers at stratospheric prices. However, the architectural details allow software teams to begin optimizing for the next-generation hardware before it is widely available.

NVIDIA’s fiscal Q2 2025 results, released in late August, report $30 billion in revenue, a 122% year-over-year increase. Data center GPU demand drives it, from cloud providers, AI startups, and enterprises building training infrastructure. Moreover, the company’s gross margin exceeds 75%. These numbers validate the thesis that the AI infrastructure buildout is a multi-year capital expenditure cycle. Its magnitude is comparable to the build-out of cloud infrastructure in the 2010s. NVIDIA’s market capitalization briefly crosses $3 trillion, making it the most valuable public company in the world by that metric. As a result, GPU manufacturing is confirmed as the most strategically important hardware segment in the global technology industry.

3.9 Apple – iOS 18 and Apple Intelligence

Apple’s iOS 18 begins shipping to consumers in September 2024. It brings the first wave of Apple Intelligence features to iPhone 15 Pro and iPhone 15 Pro Max. In practice, the initial release is modest by AI announcement standards, covering Writing Tools, photo cleanup, and smart reply suggestions. The architecture Apple describes at WWDC in June is nevertheless ambitious. It combines a 3-billion-parameter on-device language model for privacy-sensitive tasks with cloud inference routed through Apple’s Private Compute Cloud, under cryptographic privacy guarantees. In addition, an integration with ChatGPT for requests that exceed on-device capability. The ChatGPT integration signals that Apple’s AI strategy involves partnerships rather than full model ownership. That choice has significant implications for user privacy and for competitive dynamics in the mobile AI market, where Apple controls the dominant premium device ecosystem.

The Apple Intelligence roll-out is regionally staged, with EU availability delayed due to regulatory uncertainty around the Digital Markets Act. This creates the first major case in which EU regulation directly constrains Apple’s distribution practices in a way that affects consumers outside the EU. As such, it previews the compliance complexity of the next phase of AI product development, in globally-distributed software ecosystems operating under multiple overlapping regulatory frameworks.

4. AI & Technology Impact – A Season of Inflection Points

The AI landscape in Q3 2024 is defined by convergence across three axes that had previously evolved independently. They are frontier model capability, open-source accessibility, and production infrastructure maturity. The quarter is the first in which all three axes cross thresholds simultaneously, creating a compound effect greater than the sum of its parts. Therefore, understanding the implications requires tracing each axis separately before connecting them.

4.1 Q3 2024 timeline: August

4.2 Q3 2024 timeline: September

4.3 The Open-Source versus Closed-Source Inflection

The philosophical tension between Meta’s open-weights strategy and OpenAI’s closed API model is not merely commercial. Rather, it carries implications for AI safety, competitive dynamics, regulatory compliance, and the future distribution of AI’s economic value. Proponents of open weights argue that wider scrutiny improves safety through red-teaming by the entire research community. They add that enterprises must own their own model weights to meet data governance requirements. Finally, they warn about the current concentration of frontier AI capability in three or four companies. It creates dangerous single points of failure in the global information infrastructure. Proponents of controlled access argue that open-weights models can be fine-tuned to remove safety guardrails, and that responsible deployment of frontier capability requires vendor accountability. In addition, they hold that the cost of training frontier models is simply too high for open-source governance structures to sustain over multiple generations.

Q3 2024 does not resolve this debate, but it sharpens it considerably. The release of Llama 3.1 405B demonstrates empirically that Meta can train a GPT-4-class model and give it away without apparent commercial disadvantage. Its downstream businesses in social media, advertising, and virtual reality benefit from the ecosystem effects. The o1 announcement demonstrates simultaneously that the frontier continues to advance beyond what open-weights releases capture: reasoning-trained models are qualitatively different from next-token-prediction models. In addition, the reinforcement learning infrastructure required to train them is a separate and significant technical challenge not yet replicated in the open-source community. In short, both facts are true simultaneously, and the industry internalizes both.

4.4 Coding AI Matures Into Production Tooling

GitHub Copilot is now available in multiple tiers and integrated into GitHub.com alongside VS Code, JetBrains IDEs, and the command line. It begins adding multi-model support during the second half of 2024. Copilot requests can now route to different underlying models: GPT-4o for complex reasoning, GPT-4o mini for fast completions, and eventually third-party models. That reflects a growing enterprise requirement for model selection transparency and cost control in AI development tooling. GitHub Copilot Workspace was announced in preview in April 2024. In practice, it allows developers to describe a task in natural language and receive a full implementation plan with code changes. That is the beginning of an agentic coding workflow within the existing GitHub pull-request paradigm. It demonstrates Microsoft’s intention to make AI the default experience layer of the entire software development lifecycle.

Cursor, an AI-native code editor built on the VS Code foundation, accumulates a rapidly growing user base during Q3 2024. The editor integrates model selection, inline editing, codebase-wide context retrieval, and multi-file editing into a single coherent workflow. In fact, many developers find it superior to the Copilot plugin model. Consequently, by the end of the quarter, Cursor’s waiting list grows to tens of thousands of developers. In addition, its team pricing tier begins appearing in engineering team budget discussions that previously only included GitHub Copilot as the AI-assisted coding option. The emergence of credible alternatives to GitHub’s dominance in developer AI tooling is a significant structural development. It accelerates considerably across subsequent quarters as the agentic coding category expands.

5. Key Voices & Thought Leaders

Q3 2024 produces a rich body of public technical commentary from researchers, engineers, and executives. The voices that shape how the industry interprets the quarter’s events span a wide spectrum, from Meta’s C-suite to independent bloggers. Collectively they influence tens of thousands of working engineers making day-to-day decisions about AI tool selection and architectural approach.

5.1 Mark Zuckerberg – The Strategic Case for Open Source AI

Zuckerberg’s public advocacy for open-source AI in Q3 2024 is unusually direct for a CEO whose company has a commercial stake in the outcome. A Dwarkesh Patel podcast interview follows on July 24, the day after the Llama 3.1 release. In it he addresses the economics of the open-weights bet in detail. Meta benefits from an open ecosystem because it makes the Llama ecosystem larger and reduces the cost of running Meta AI on its platforms. In addition, it attracts AI researchers who want to work on models that have broad societal impact.

He is candid about the training cost of the 405B model and about the tradeoffs of releasing it freely. It is one of the most substantive public discussions of frontier AI economics by any industry figure in 2024. The interview is widely shared and becomes a reference point for discussions about AI governance and competitive strategy throughout the industry.

5.2 Simon Willison – Prolific Technical Chronicler

Simon Willison, co-creator of Django and author of one of the most widely read AI engineering blogs, publishes extensively throughout Q3 2024.

His posts on Llama 3.1 provide unusually detailed practical guidance. They cover how to quantize and run the 70B model on a Mac Studio, and how the 128K context window changes retrieval architectures. Finally, they also explain how the community license restricts use cases that practitioners might not have considered. Willison’s distinctive contribution to technical discourse is his commitment to actually running the software he writes about. In addition, his hands-on notes on llama.cpp, Ollama, and the llm CLI tool are used as tutorials by thousands of developers. His documentation of the CrowdStrike incident’s technical root cause is also among the most widely shared post-mortems of the event. It combines rigorous analysis of the kernel-mode driver failure with clear explanations accessible to developers who do not specialize in systems programming.

5.3 Andrej Karpathy – The AI Educator’s Perspective

Former OpenAI co-founder and Tesla AI director Andrej Karpathy remains one of the most followed voices on AI architecture and education in Q3 2024. He posts on the significance of Llama 3.1. His focus is how Chinchilla scaling laws apply when a model trains on substantially more tokens than the “compute-optimal” amount. As a result, the posts contribute to public understanding of why smaller, more data-efficient models are becoming increasingly competitive with larger ones trained for shorter periods. Karpathy emphasises building intuition about neural network behavior rather than treating models as black-box APIs. That influences a generation of developers engaging with the open-source ecosystem in Q3 2024. It also contributes to a culture of model evaluation and understanding that goes beyond simple benchmark comparison.

5.4 Sam Altman – Framing the Reasoning Model Transition

OpenAI CEO Sam Altman’s communication around the o1 release is notably measured compared to previous OpenAI product announcements. The official messaging does not claim capability superiority across all tasks. Instead, it positions o1 explicitly as a tool for hard reasoning problems. It also acknowledges that GPT-4o remains faster and more cost-effective for the majority of use cases. This positioning is a product hierarchy rather than a single best model. It reflects a maturing understanding within OpenAI of how AI models fit into complex real-world workflows. In addition, it influences how the broader industry begins to think about model selection. It becomes a product design decision requiring explicit tradeoff analysis. The simple “use the best available” heuristic made sense when there was only one frontier model family.

5.5 Yann LeCun – The Open-Source AI Advocate in Chief

Meta’s Chief AI Scientist Yann LeCun continues in Q3 2024 to advocate for open-source AI. He is the most vocal and intellectually rigorous voice on the subject among senior industry figures. His position is that the current generation of autoregressive large language models is a necessary but insufficient path to general AI. In addition, he holds that open research is the only way to discover what replaces them. The position provides a philosophical grounding for Meta’s open-weights strategy that extends beyond commercial interests.

LeCun’s social media commentary is widely shared and frequently debated. It covers the Llama 3.1 benchmarks and the limitations of current reinforcement learning approaches to alignment. In addition, it takes up the EU AI Act’s treatment of general-purpose models. He argues consistently that concentration of frontier AI in a small number of closed-source companies poses greater societal risks than open-weights models. The position remains contested but influential in the academic and policy communities that will shape AI governance over the following years.

6. Trend Synthesis – The Architecture of a New AI Market

Looking across the events of Q3 2024, three structural shifts are visible. They extend beyond the individual announcements and define the shape of the market forming around AI technology. They also cover the engineering disciplines required to build within it.

The first shift is the bifurcation of the AI model market into a cost-performance tier and a capability-research tier. GPT-4o mini, Llama 3.1 8B, and a growing set of smaller fine-tuned models occupy the cost-performance tier. They are fast, cheap, and capable enough for the vast majority of production workloads. Moreover, they are increasingly commoditized as multiple providers offer similar quality at similar prices.

GPT-4o, Claude 3.5 Sonnet, and Llama 3.1 405B occupy a mid-capability tier suitable for complex reasoning and nuanced generation. OpenAI o1, still in preview, occupies a nascent capability-research tier for tasks that demand deliberate multi-step reasoning rather than rapid pattern completion. This tiered structure parallels the cloud compute market’s evolution from single instance pricing to a diverse menu of instance types optimized for different workload profiles. In addition, it signals a change in AI product design. Explicit choices about model tier, cost tolerance, and latency requirements increasingly replace a single default selection.

6.1 AI infrastructure becomes critical infrastructure

The second shift is the emergence of AI infrastructure as critical infrastructure. The regulatory and operational sense is the same as for financial systems, telecommunications, and energy grids. CrowdStrike makes this concrete. For example, a software component running with kernel-level privileges is auto-updated at scale. It causes more immediate disruption to global economic activity than most cyberattacks on record.

The insurance industry responds with detailed coverage analyses and updated underwriting criteria for software resilience risk. In addition, preliminary policy discussions cover mandatory testing regimes for kernel-level software. It is the beginning of a long process. Software governance frameworks must catch up to the operational reality that software is now civilization’s load-bearing infrastructure. AI systems are being embedded into healthcare diagnostics, financial risk models, and critical infrastructure monitoring at an accelerating rate. Consequently, they will face analogous governance frameworks in the years following Q3 2024. The CrowdStrike incident is not a warning about AI specifically. However, it is an extremely clear demonstration of what happens when software governance lags behind deployment scale.

6.2 The open-source AI tooling ecosystem matures

The third shift is the maturation of the open-source AI tooling ecosystem from experimental libraries into production-grade infrastructure. The stack combines llama.cpp for inference, Ollama for local deployment, and LiteLLM for multi-provider routing. In addition, GraphRAG adds structured knowledge retrieval, and LangChain and LlamaIndex handle application orchestration. It gives engineering teams a coherent, largely free, and increasingly reliable way to build AI-powered applications without proprietary cloud dependencies. The stack is not yet as mature as the Kubernetes ecosystem that reached production readiness around 2017–2018, but the trajectory is clear. By the end of Q3 2024, the default question in AI application architecture is no longer whether to use open-source tooling. It is how to assemble the right combination of open-source components for a given use case. A second question is how to balance the operational advantages of managed cloud services against the data governance advantages of self-hosted deployments.

6.3 How the three shifts compound

These three shifts interact in ways that compound their individual effects. A bifurcated model market rewards cost optimization expertise alongside model evaluation skills, creating new specialization paths in AI engineering. The emergence of AI as critical infrastructure changes the status of governance tooling. Audit logging, access control, usage monitoring, and prompt injection detection become first-class engineering concerns rather than compliance afterthoughts. The maturity of open-source AI infrastructure lets smaller organizations participate in the frontier at a fraction of the cost required twelve months earlier. That democratises AI capability and accelerates adoption across industry sectors which previously could not justify the expense. Together, these dynamics define Q3 2024. Overall, AI completes its transition from a feature that differentiates products to a platform that underlies them.

7. Summary

Q3 2024 is a quarter of paradigm shifts arriving in rapid succession. Meta’s Llama 3.1 405B release on July 23 is the most significant event in the history of open-source AI. For the first time, an organization can download, fine-tune, and deploy a model that competes on benchmark performance with OpenAI’s best closed-source offering. Moreover, it runs on their own hardware, with no API costs and no data leaving their infrastructure. This changes the strategic calculus for every enterprise that was previously locked into cloud-only AI procurement. The absence of viable open-weights alternatives at frontier capability levels no longer binds them. Llama 3.1 combines a 128K context window, 15T token training, and genuinely permissive commercial licensing. As such, it is not a research curiosity but a production-grade option for the majority of enterprise AI workloads.

OpenAI’s o1 preview release on September 12 demonstrates simultaneously that the frontier continues to advance along dimensions that open-weights releases have not yet reached. A reasoning model thinks through problems with an internal chain of thought. It achieves PhD-level performance on scientific benchmarks and near-expert performance on competitive mathematics. That is qualitatively different from a larger version of the same autoregressive architecture. The message is not that closed-source AI is better across all dimensions. Rather, it is that the diversity of AI model architectures is increasing. No single model or approach meets all use cases across the full spectrum of latency, cost, accuracy, and domain-specialization requirements that real-world applications impose.

7.1 CrowdStrike, the AI Act, and what to watch

The CrowdStrike outage on July 19 is not an AI story. However, it is the most important infrastructure story of the quarter and arguably of the year. Its relevance to AI is prospective: AI systems are being embedded into critical infrastructure at an accelerating rate. In addition, the lesson of CrowdStrike concerns kernel-level or deep integration into operating systems and critical processes. It demands governance, staged rollout, and rollback capability that the current software industry has not yet fully standardized. Therefore, those standards should be built proactively, before an AI-specific incident of equivalent scale. It is the technical policy challenge that emerges most clearly from Q3 2024. Anyone thinking seriously about the long-run trajectory of AI deployment in high-stakes environments faces it.

The EU AI Act’s entry into force on August 1 is the opening movement of a multi-year compliance transformation. Its immediate practical effect is modest; its structural effect is large. In practice, every AI-powered product serving European users is now subject to a risk-based regulatory framework. In addition, the compliance work beginning in Q3 2024 defines organizational AI governance practices that will outlast the specific provisions of the current text. Technical maturity, economic scale, and regulatory attention all characterize Q3 2024. The combination marks the moment when AI completes its transition. A research initiative running alongside the technology industry becomes an industry sector in its own right. It carries the full complement of governance expectations, infrastructure obligations, and market dynamics that characterize any sector of civilization-scale importance.

8. Sources

1. OpenAI – Introducing GPT-4o mini: advancing cost-efficient intelligence (July 18, 2024)
2. Meta AI – Introducing Llama 3.1: Our most capable models to date (July 23, 2024)
3. Dubey et al. – The Llama 3 Herd of Models, arXiv:2407.21783 (July 31, 2024)
4. Dwarkesh Patel Podcast – Mark Zuckerberg: Llama 3, Open Sourcing $10B Models (July 24, 2024)
5. CrowdStrike – Falcon Content Update: Remediation and Guidance Hub (July 2024)
6. Microsoft – Helping our customers through the CrowdStrike outage (July 20, 2024)
7. European Union – Regulation (EU) 2024/1689: The Artificial Intelligence Act, Official Journal (July 12, 2024)
8. OpenAI – o1 System Card (September 12, 2024)
9. OpenAI – Learning to Reason with LLMs (September 12, 2024)
10. Meta AI – Llama 3.2: Revolutionizing edge AI and vision (September 25, 2024)
11. OpenAI – GPT-4o fine-tuning is now available (August 20, 2024)
12. OpenAI – ChatGPT Advanced Voice Mode update (September 24, 2024)
13. Microsoft Research – GraphRAG: Graph-based Retrieval-Augmented Generation (2024)
14. Georgi Gerganov – llama.cpp: LLM inference in C/C++ (2023–2024)
15. Ollama – Get up and running with large language models (2024)
16. BerriAI – LiteLLM: Call all LLM APIs using the OpenAI format (2024)
17. Simon Willison – Llama 3.1 405B, 70B and 8B (July 23, 2024)
18. The Verge – IT outage takes down banks, airlines, and emergency services worldwide (July 19, 2024)
19. The Verge – Meta releases its first open AI model that can process images (September 25, 2024)
20. TechCrunch – Meta’s Llama AI models get multimodal (September 25, 2024)
21. NVIDIA – GB200 NVL72 Data Center Platform (2024)
22. Apple – Apple Intelligence overview (2024)
23. GitHub – GitHub Copilot Workspace: Welcome to the beginning of a new era for developer AI (April 2024)

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.