Chronological feed of everything captured from François Chollet.
This tutorial details the fine-tuning of the Gemma model on TPU v5 hardware. It highlights a toolchain consisting of Kinetic, Keras, and JAX, presented as an optimized stack for leveraging TPUs at scale. The associated script further elaborates on setups, technical specifics, and practical considerations of using Kinetic.
JAX represents a well-designed low-level machine learning framework. Its design principles facilitate superior performance with reduced development effort. Conversely, poorly designed frameworks hinder performance and increase effort.
A recent tutorial demonstrates fine-tuning large language models (LLMs) using Keras Kinetic, an extension for Keras facilitating model training with JAX and Cloud TPUs. This approach is exemplified by fine-tuning the Gemma 2B model on the PubMedQA dataset, indicating potential for efficient medical question-answering system development.
Keras Kinetic introduces a streamlined approach to remote execution of machine learning workloads on cloud TPUs and GPUs. It automates containerization, dependency management, and deployment to GKE clusters, simplifying the transition from local development to scalable cloud execution. This allows developers to run functions on powerful accelerators with minimal configuration overhead.
AI integration presents a significant opportunity for established companies with existing profitable business models. By leveraging AI to enhance current offerings and develop new, AI-first products, these companies can solidify their market position and drive further growth. This strategy is exemplified by products like Adobe Podcast, which demonstrates the potential for AI to both improve and innovate within an established company.
François Chollet, a prominent AI researcher, identified Adobe Podcast as a top-tier AI product. This endorsement highlights the effective application of AI within the audio editing domain, suggesting that the product demonstrably leverages AI to deliver a superior user experience or functionality.
While OpenClaw demonstrated the product-market fit for local AI assistants, its lack of security architecture limited production deployment. PokeeClaw addresses these vulnerabilities by implementing a sandboxed environment featuring RBAC, approval workflows, and audit trails to enable enterprise-safe agentic workflows.
Human intelligence, amplified by externalized cognitive infrastructure like computers and the internet, can rapidly achieve expert-level performance in complex, rule-based systems. An experiment involving learning chess ("Glurg") rules from scratch demonstrates that a 3000 Elo engine could be developed within 24 hours, and a 3500 Elo engine with significantly improved efficiency within three weeks. This suggests human intelligence is near-optimal in its ability to quickly master rule-based domains.
This content redefines intelligence not as an unbounded scalar but as a conversion ratio with an optimality bound, akin to making a ball rounder rather than a tower taller. It posits that while individual humans may not be optimally intelligent, a collective of intelligent humans augmented by external tools approaches this bound. The author argues that humanity’s ability to solve problems is near-optimal given available information, with current AI amplifying this collective intelligence.
The advent of Artificial General Intelligence (AGI) is projected to redefine societal stratification, shifting the basis of class division from material wealth to cognitive agency. This future societal structure will delineate between individuals who maintain control over their attention and actions (the "focus class") and those whose reward mechanisms are entirely managed by AI systems (the "slop class"). This division implies a fundamental change in how individuals interact with and are influenced by advanced AI.
François Chollet, creator of Keras and the ARC AGI benchmark, discusses NDIA, a new AI research lab focused on symbolic program synthesis as an alternative to deep learning. NDIA aims to build AI that requires less data, runs more efficiently, and generalizes better by replacing parametric curves with concise symbolic models, addressing the limitations of current LLM-based approaches. This new approach, which aims for optimal AI by leveraging symbolic models, is driven by the belief that current deep learning methods, while effective for verifiable domains, are inefficient and will not lead to true AGI.
To generalize an AI system beyond a specific task (ARC-AGI-3), it is necessary to remove all components engineered or configured based on test runs on those specific tasks. This primarily includes prompts detailing the process to solve the games.
François Chollet announced that ARC-AGI-4 is slated for an early 2027 release, initiating an annual benchmark release cycle. Each new benchmark aims to be "fully unsaturated upon release" and address "the most important unanswered research questions." This development strategy necessitates anticipating future AI capabilities during the benchmark design phase, echoing the approach taken for ARC-AGI-3.
François Chollet clarifies his long-standing definition of Artificial General Intelligence (AGI), emphasizing learning efficiency over task-specific performance benchmarks. He posits that AGI should autonomously master any human-learnable task with equivalent learning efficiency, diverging from current AI development that often targets specific capabilities. This reorientation shifts the focus from achieving a pre-defined "target" to developing a "compass" for continuous, human-like learning.
The ARC-AGI-3 benchmark evaluates AI agentic intelligence through interactive reasoning environments that require human-level action efficiency on novel tasks without prior training. This benchmark highlights a significant gap between current frontier AI models, which perform under 1%, and human ability, as humans can solve all tasks upon first contact. The competition offers public environments for testing and private test sets for evaluation, aiming to drive advancements in general artificial intelligence.
ARC-3 emphasizes interactive learning, goal discovery, and temporal planning in novel environments. It aims to measure efficient skill acquisition, a defining characteristic of general intelligence, by scaling up these capabilities within a "micro-AGI" framework, rather than focusing on perception or data-driven approaches like LLMs.
This GitHub repository offers Jupyter notebooks complementing the "Deep Learning with Python, third edition" by Chollet and Watson. It provides runnable code samples for practical application of theoretical concepts. The notebooks are designed for use with Google Colab, leveraging its free GPU runtime, and support Keras 3 with JAX, TensorFlow, or PyTorch backends. Users should refer to the companion book for comprehensive understanding, as the notebooks intentionally omit explanatory text and figures.
François Chollet, creator of Keras, argues that LLMs are fundamentally pattern-memorization systems — "databases of vector programs" — that can only operate within their training data distribution, making them categorically distinct from general intelligence. He defines intelligence as the efficiency with which an agent acquires new skills in novel, unprepared-for situations (operationalized via his ARC benchmark), and contends that LLMs score near zero on this metric. Chollet traces the failure mode to the architecture itself: transformers excel at passive, Hebbian-style associative learning but lack the active, causal, experimental learning that characterizes human cognition. While LLMs are practically valuable for automating tasks within known distributions, existential risk narratives are unfounded — the real bottleneck to AGI is unsolved program synthesis and few-shot generalization, not scaling.
François Chollet argues that the pre-training scaling paradigm fundamentally cannot produce general fluid intelligence because LLMs only acquire static, memorized skills — not the ability to synthesize novel solutions on the fly. Test-time adaptation (TTA) is a meaningful step forward, but remains compute-inefficient and lacks compositional generalization. True AGI, in Chollet's framing, requires combining two forms of abstraction: value-centric (continuous, perception/intuition via deep learning) and program-centric (discrete, reasoning via combinatorial search), and his new lab Ndea is building a deep learning-guided program search system targeting exactly this hybrid architecture.
Namex is a Python utility designed to strictly separate a package's implementation from its public API. It enables developers to define an explicit allowlist of public symbols, offering precise control over visibility, naming, and exposure paths. This facilitates easier refactoring, prevents accidental exposure of private utilities, and simplifies API version control.
François Chollet, creator of Keras and AI researcher at Google, challenges the common "intelligence explosion" narrative, arguing that intelligence is not an isolated property but emerges from interaction between a brain, body, and environment. He posits that focusing solely on brain (or algorithm) improvements ignores crucial bottlenecks and external dependencies, leading to an oversimplified view of AI progress. Chollet suggests that general AI systems, like science itself, will face exponential friction, leading to linear, not exponential, overall progress despite increasing resource consumption.
This Python Gist demonstrates a method for creating a unified interface for numerical operations that can seamlessly handle both NumPy arrays and Keras backend tensors (e.g., TensorFlow). It achieves this by dynamically dispatching calls to either the NumPy implementation or the Keras backend implementation based on the input type. This enables writing code once that can operate efficiently with different numerical computing frameworks.
Keras functional API supports blending imperative ops like tf.exp and constant tensors into symbolic layer graphs, but encounters runtime errors in eager execution for complex recurrent models. In seq2seq LSTMs, passing encoder states as initial_state to decoder LSTM triggers ValueError during RNN step computation. The failure stems from unhandled DeferredTensor objects in matmul operations within the LSTM cell, blocking tensor conversion.
François Chollet shares introspective notes on software engineering practices drawn from his experience. Key principles include prioritizing simplicity, modularity, and testability to enhance code reliability and maintainability. The post emphasizes disciplined habits like writing tests first and avoiding over-engineering for sustainable productivity.
The provided content contains only the title and metadata of François Chollet's blog post "What Worries Me About AI," with no substantive body text ingested. Core insights on AI worries cannot be extracted due to absence of detailed arguments or claims. Analysis is constrained to surface-level identification of the topic as expressing concerns about AI from a prominent AI researcher.
Hualos is a demo project using a Flask server with gevent to expose an API for publishing and consuming JSON training events from Keras' RemoteMonitor callback. The landing page at localhost:9000 consumes these events and renders metrics in real-time using c3.js graphs built on d3.js. Integration requires starting the server with api.py, loading the page, and adding RemoteMonitor(root='http://localhost:9000') to model.fit callbacks.
François Chollet argues that intelligence explosion—recursive self-improvement leading to superintelligence—is implausible due to fundamental limits in generalization from finite data. Intelligence is defined by adapting to novel situations via compression of prior knowledge, not raw optimization power. Scaling compute and data cannot overcome the combinatorial explosion of possible environments, making ASI unreachable through brute-force methods.
API design prioritizes user experience through three rules: deliberately designing end-to-end workflows that map to domain concepts without exposing implementation details; reducing cognitive load via consistent naming, minimal new concepts, balanced parameterization, automation, and example-rich docs; and providing interactive feedback with early error catching, detailed actionable messages, and user support channels. A litmus test for quality is whether users can recall common workflows without docs after one exposure. These principles derive from empathizing with all users, countering smart engineer syndrome and masochistic attitudes toward complexity.
Keras introduces a new RNN API using RNN(cells) for stacking LSTM layers, achieving 15% faster training than sequential LSTM layers on CPU. Benchmark on 10k samples of 60 timesteps and 64 dims shows classic stacked LSTMs at 35s/epoch versus 30s/epoch for the new approach. Both use RMSprop and MSE loss with batch size 128 over 4 epochs.
Deep learning will evolve from pure differentiable geometric transformations to program-like models blending algorithmic primitives (e.g., loops, conditionals, data structures) with neural layers, enabling reasoning, abstraction, and extreme generalization beyond current pattern recognition limits. Training will shift beyond backpropagation to non-differentiable methods like genetic algorithms and evolution strategies, paired with automated architecture search (AutoML) and lifelong learning via reusable modular subroutines from a global meta-learning library. This enables efficient model growth with minimal human engineering, achieving human-like generalization across tasks using sparse new data.
Deep learning models perform continuous geometric transformations on high-dimensional vector spaces, effectively mapping input manifolds to output manifolds given dense training data. However, they cannot represent discrete reasoning, long-term planning, or algorithmic tasks like generating code from specifications or learning sorting algorithms, regardless of data scale. They achieve local generalization near training data but lack human-like extreme generalization for novel situations, remaining brittle to adversarial perturbations without true causal understanding.
François Chollet provides a downsized Xception CNN architecture omitting residual connections, tailored for 200x200x3 inputs and 100-way classification. It employs an initial 3x3 Conv2D(32, stride=2) with ReLU and max pooling, followed by three depthwise-separable Conv2D blocks (128, 256, 512 filters) each with dual 3x3 SeparableConv2D layers, BatchNorm, ReLU, and stride-2 pooling. The model culminates in global average pooling and softmax output, prioritizing efficiency via separable convolutions.
François Chollet's Gist provides minimal Keras examples for 1D MSE linear regression using a single Dense layer. It extends to binary logistic regression with sigmoid activation and binary_crossentropy loss. A third variant incorporates L1/L2 regularization via l1l2 on the weight matrix.
Deep learning has advanced rapidly to near-human performance in tasks like speech/image recognition and Go, yet remains underexploited in everyday products and processes. Analogous to the Internet's eventual ubiquity, AI will permeate all industries, automating intellectual tasks, disrupting jobs, and enabling a prosperity era—but only if made accessible to non-experts. Keras lowers barriers by simplifying deep learning for users with basic CS literacy, fostering widespread value creation as demonstrated by startups like Comma.ai; early adopters must prioritize open tools, tutorials, and knowledge sharing to prevent elite capture and ensure positive outcomes.
François Chollet's Keras script demonstrates fine-tuning VGG16 on a small cats-vs-dogs dataset by freezing the first 25 layers, adding a custom binary classifier on top, and using heavy data augmentation with SGD at low learning rate. The approach leverages ImageNet pretraining for convolutional base while training only top layers on 2000 training and 800 validation images of 150x150 pixels over 50 epochs. Key hyperparameters include batch size 16, momentum 0.9, and augmentation via shear, zoom, and flips to combat overfitting.
François Chollet's Keras script demonstrates transfer learning by extracting VGG16 bottleneck features from 2000 training and 800 validation images (1000/400 cats and dogs each), saving them as NumPy arrays, then training a simple top classifier (Flatten-Dense256-Dropout-Dense1 sigmoid) with RMSprop and binary crossentropy for 50 epochs. Common issues include using 'wb' mode for np.save/np.load to avoid UnicodeDecodeError, understanding bottleneck_features as (N, 4, 4, 512) feature maps rather than probabilities, and adapting for multi-class via softmax/categorical_crossentropy. Prediction code uses VGG16 for new image features fed to the top model or full fine-tuned model generators with argmax on class indices.
François Chollet's Keras script demonstrates building a CNN for binary image classification (cats vs. dogs) using only 2000 training images (1000 per class) and 800 validation images. Key technique is heavy data augmentation during training (shear, zoom, flips) with a simple 3-layer Conv2D architecture trained for 50 epochs on 150x150 RGB images. Model compiles with binary crossentropy and RMSprop, saving weights post-training, enabling strong generalization on limited data.
François Chollet proposes a functional Keras API where layers are callable on input tensors, enabling concise graph model construction via tensor chaining and topology tracking. Key features include shared layers via reuse, Lambda for arbitrary ops, merge functions, and backward-compatible Model compilation/training with flexible input/output dicts. Discussions resolve masking via node propagation, layer querying for weight transfer, and Sequential integration by making models callable.
François Chollet demonstrates defining a Theano function to compute and output activations from intermediate layers in a Keras Sequential model. The approach uses model.layers to access layer inputs and outputs, creating a function like theano.function([model.layers[0].input], model.layers[1].output(train=False)). This allows direct transformation of input batches through specific layers without full forward passes, useful for visualization and analysis in Theano-backed Keras.
Scientific and technological progress, despite exponential increases in resources like researchers and computing power, generally proceeds at a linear rate. This is because the difficulty of making impactful discoveries within a given field increases exponentially over time, effectively canceling out the benefits of increased resources. Therefore, the notion of an "intelligence explosion" or technological "Singularity" driven by exponential progress is fundamentally flawed; even a self-improving AI would face this linearity constraint without exponentially increasing resources.
Current web platforms prioritize information flow and commercial interests, leading to "collective stupidity" and a focus on low-quality, attention-grabbing content. There is a critical need to redesign these platforms to incorporate psychological aspects of content creation and consumption, fostering higher quality content, genuine creativity, and collective intelligence. This involves shifting from content-neutral models to those that actively shape and improve the quality of user-generated content by focusing on project-driven engagement, motivational feedback, curated inspiration, and accessible learning.
The internet, as currently structured, primarily fosters "collective stupidity" rather than "collective intelligence." This is due to an infrastructure that prioritizes attention-grabbing, "fun" content, epitomized by "piano-playing-cat" videos, over meaningful, productive interactions. This paradigm, driven by view-count-based popularity, leads to a significant waste of human time and potential, as evidenced by the billions of hours spent on platforms like Facebook with little to no genuine return for users.