flâneur

Tony Kam

1 followers · 2 following · 1330 views

on the atlas — 47

highlights — 264

  • And it turns out that price-sensitive customers are often significantly worse risks; there’s often a reason why someone’s so keen on getting the cheapest deal, and it’s usually a reason that isn’t in the database you bought.
    A Cynic’s Guide To Fintech. Several business models that are bound… | by Dan Davies | Bull Market | Medium
  • Surprisingly, Nano Banana is terrible at style transfer even with prompt engineering shenanigans, which is not the case with any other modern image editing model. I suspect that the autoregressive properties that allow Nano Banana’s excellent text editing make it too resistant to changing styles. That said, creating a new image in the style of Studio Ghibli does in fact work as expected, and creating a new image using the character provided in the input image with the specified style (as opposed to a style transfer) has occasional success.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • That’s honestly better than expected, and the prompt cost 916 tokens. It got the overall layout and colors correct: the issues are more in the text typography, leaked classes/styles/JavaScript variables, and the sidebar:body ratio. No, there’s no practical use for having a generative AI render a webpage, but it’s a fun demo.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • my guess is that this rule is a hack intended to avoid the perception of model collapse by avoiding the generation of 2022-era AI images which would be annotated with those buzzwords.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • Nano Banana is still bad at rendering text perfectly/without typos as most image generation models. However, the expanded text is peculiar: it does follow from the prompt, although “Blue Blur” is a nickname for the normal Sonic the Hedgehog. How does an image generating model generate logical text unprompted anyways?
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • One of the most compelling-but-underdiscussed use cases of modern image generation models is being able to put the subject of an input image into another scene. For open-weights image generation models, it’s possible to “train” the models to learn a specific subject or person even if they are not notable enough to be in the original training dataset using a technique such as finetuning the model with a LoRA using only a few sample images of your desired subject. Training a LoRA is not only very computationally intensive/expensive, but it also requires care and precision and is not guaranteed t…
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • Working with the Gemini API is a pain and requires annoying image encoding/decoding boilerplate, so I wrote and open-sourced a Python package: gemimg, a lightweight wrapper around Gemini API’s Nano Banana endpoint that lets you generate images with a simple prompt, in addition to handling cases such as image input along with text prompts.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • For developers who want to build apps that programmatically generate images from Nano Banana, Google offers the gemini-2.5-flash-image endpoint on the Gemini API. Each image generated costs roughly $0.04/image for a 1 megapixel image (e.g. 1024x1024 if a 1:1 square): on par with most modern popular diffusion models despite being autoregressive, and much cheaper than gpt-image-1’s $0.17/image.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • While most image generation models are diffusion-based to reduce the amount of compute needed to train and generate from such models, gpt-image-1 works by generating tokens in the same way that ChatGPT generates the next token, then decoding them into an image.
    Nano Banana can be prompt engineered for extremely nuanced AI image generation | Max Woolf's Blog
  • Context rot, by Workaccount2 on Hacker News, for the thing where model output quality falls as the context grows longer during a session.
    2025: The year in LLMs
  • The problem is that the big cloud models got better too—including those open weight models that, while freely available, were far too large (100B+) to run on my laptop. Coding agents changed everything for me. Systems like Claude Code need more than a great model—they need a reasoning model that can perform reliable tool calling invocations dozens if not hundreds of times over a constantly expanding context window. I have yet to try a local model that handles Bash tool calls reliably enough for me to trust that model to operate a coding agent on my device.
    2025: The year in LLMs
  • I remain deeply concerned about the safety implications of these new tools. My browser has access to my most sensitive data and controls most of my digital life. A prompt injection attack against a browsing agent that can exfiltrate or modify that data is a terrifying prospect.
    2025: The year in LLMs
  • Then in November Anthropic published Code execution with MCP: Building more efficient agents—describing a way to have coding agents generate code to call MCPs in a way that avoided much of the context overhead from the original specification.
    2025: The year in LLMs
  • Anthropic themselves appeared to acknowledge this later in the year with their release of the brilliant Skills mechanism—see my October post Claude Skills are awesome, maybe a bigger deal than MCP. MCP involves web servers and complex JSON payloads. A Skill is a Markdown file in a folder, optionally accompanied by some executable scripts.
    2025: The year in LLMs
  • Since leaning heavily into Claude Code and friends myself I’ve hardly used MCP at all—I’ve found CLI tools like gh and libraries like Playwright to be better alternatives to the GitHub and Playwright MCPs.
    2025: The year in LLMs
  • The reason I think MCP may be a one-year wonder is the stratospheric growth of coding agents. It appears that the best possible tool for any situation is Bash—if your agent can run arbitrary shell commands, it can do anything that can be done by typing commands into a terminal.
    2025: The year in LLMs
  • MCP’s release coincided with the models finally getting good and reliable at tool-calling, to the point that a lot of people appear to have confused MCP support as a pre-requisite for a model to use tools.
    2025: The year in LLMs
  • They also shipped Gemini CLI (their open source command-line coding agent, since forked by Qwen for Qwen Code), Jules (their asynchronous coding agent), constant improvements to AI Studio, the Nano Banana image models, Veo 3 for video generation, the promising Gemma 3 family of open weight models and a stream of smaller features.
    2025: The year in LLMs
  • Gemini 2.5 and then Gemini 3.0—each model family supporting audio/video/image/text input of 1,000,000+ tokens, priced competitively and proving more capable than the last.
    2025: The year in LLMs
  • internal knowledge and token-based reasoning capabilities.
    2025: The year in LLMs
  • In November Google fully embraced the “Nano Banana” name with the release of Nano Banana Pro. This one doesn’t just generate text, it can output genuinely useful detailed infographics and other text and information-heavy images. It’s now a professional-grade tool.
    2025: The year in LLMs
  • Nano Banana caught people’s attention because it could generate useful text! It was also clearly the best model at following image editing instructions.
    2025: The year in LLMs
  • Anyone who’s tried running their agent with automatic confirmation (aka YOLO mode—Codex CLI even aliases --dangerously-bypass-approvals-and-sandbox to --yolo) has experienced the trade-off: using an agent without the safety wheels feels like a completely different product. A big benefit of asynchronous coding agents like Claude Code for web and Codex Cloud is that they can run in YOLO mode by default, since there’s no personal computer to damage.
    2025: The year in LLMs
  • ChatGPT Code Interpreter in early 2023—a system baked into ChatGPT that allowed it to run Python code in a Kubernetes sandbox.
    2025: The year in LLMs
  • The major labs all put out their own CLI coding agents in 2025 Claude Code Codex CLI Gemini CLI Qwen Code Mistral Vibe Vendor-independent options include GitHub Copilot CLI, Amp, OpenCode, OpenHands CLI, and Pi. IDEs such as Zed, VS Code and Cursor invested a lot of effort in coding agent integration as well.
    2025: The year in LLMs
  • But if you define agents as LLM systems that can perform useful work via tool calls over multiple steps then agents are here and they are proving to be extraordinarily useful. The two breakout categories for agents have been for coding and for search.
    2025: The year in LLMs
  • But if you define agents as LLM systems that can perform useful work via tool calls over multiple steps then agents are here and they are proving to be extraordinarily useful. The two breakout categories for agents have been for coding and for search.
    2025: The year in LLMs
  • But if you define agents as LLM systems that can perform useful work via tool calls over multiple steps then agents are here and they are proving to be extraordinarily useful. The two breakout categories for agents have been for coding and for search.
    2025: The year in LLMs
  • But who is actually going to build the LLM GUI? In this world view, nano banana is a first early hint of what that might look like. And importantly, one notable aspect of it is that it's not just about the image generation itself, it's about the joint capability coming from text generation, image generation and world knowledge, all tangled up in the model weights.
    2025 LLM Year in Review | karpathy
  • We're going to see equivalents of personal computing, of microcontrollers (cognitive core), or internet (of agents)
    2025 LLM Year in Review | karpathy
  • Google Gemini Nano banana is one of the most incredible, paradigm-shifting models of 2025. In my world view, LLMs are the next major computing paradigm similar to computers of the 1970s, 80s, etc
    2025 LLM Year in Review | karpathy
  • Vibe coding will terraform software and alter job descriptions.
    2025 LLM Year in Review | karpathy
  • Note that the primary distinction that matters is not about where the "AI ops" happen to run (in the cloud, locally or whatever), but about everything else - the already-existing and booted up computer, its installation, context, data, secrets, configuration, and the low-latency interaction.
    2025 LLM Year in Review | karpathy
  • And while agent swarms running in the cloud feels like the "AGI endgame", we live in an intermediate and slow enough takeoff world of jagged capabilities that it makes more sense to run the agents directly on the developer's computer.
    2025 LLM Year in Review | karpathy
  • Claude Code (CC) emerged as the first convincing demonstration of what an LLM Agent looks like - something that in a loopy way strings together tool use and reasoning for extended problem solving. In addition, CC is notable to me in that it runs on your computer and with your private environment, data and context.
    2025 LLM Year in Review | karpathy
  • Personally I suspect that LLM labs will trend to graduate the generally capable college student, but LLM apps will organize, finetune and actually animate teams of them into deployed professionals in specific verticals by supplying private data, sensors and actuators and feedback loops.
    2025 LLM Year in Review | karpathy
  • LLM apps like Cursor bundle and orchestrate LLM calls for specific verticals: They do the "context engineering" They orchestrate multiple LLM calls under the hood strung into increasingly more complex DAGs, carefully balancing performance and cost tradeoffs. They provide an application-specific GUI for the human in the loop They offer an "autonomy slider"
    2025 LLM Year in Review | karpathy
  • we got a whole new knob (and and associated scaling law) to control capability as a function of test time compute by generating longer reasoning traces and increasing "thinking time"
    2025 LLM Year in Review | karpathy
  • "shape" of LLM intelligence in a more intuitive sense. We're not "evolving/growing animals", we are "summoning ghosts"
    2025 LLM Year in Review | karpathy
  • People who build good internal models of this new intelligent entity will be better equipped to reason about it today and predict features of it in the future.
    The space of minds | karpathy
  • The more a task/job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization fingers crossed, or via weaker means like imitation. This is what's driving the "jagged" frontier of progress in LLMs.
    Verifiability | karpathy
  • In this new programming paradigm then, the new most predictive feature to look at is verifiability. If a task/job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well.
    Verifiability | karpathy
  • A lot of recent work is also very interested in memory (think CLAUDE.md files) as a mechanism for test-time learning that uses the text/context as the substrate instead of weights.
    Animals vs Ghosts | karpathy
  • LLMs do have their own continual learning at test time, it's just not based on weight training, but I think Sutton didn't fully react to that
    Animals vs Ghosts | karpathy
  • Animals do observe demonstrations, but they are not strictly speaking directly supervised with actions, like supervised learning does. Animals are never teleoperated in training mode. The closest thing I can think of is if you for example help a child eat with a spoon or something, by literally holding their hand and showing the motion. Even then, it's not clear that their brains are literally training on that.
    Animals vs Ghosts | karpathy
  • It's possible that ghosts:animals :: planes:birds.
    Animals vs Ghosts | karpathy
  • Stated plainly, today's frontier LLM research is not about building animals. It is about summoning ghosts. You can think of ghosts as a fundamentally different kind of point in the space of possible intelligences. They are muddled by humanity. Thoroughly engineered by it. They are these imperfect replicas, a kind of statistical distillation of humanity's documents with some sprinkle on top. They are not platonically bitter lesson pilled, but they are perhaps "practically" bitter lesson pilled, at least compared to a lot of what came before.
    Animals vs Ghosts | karpathy
  • TLDR: Pretraining is our crappy evolution. It is one candidate solution to the cold start problem, to be followed later by finetuning on tasks that look more correct
    Animals vs Ghosts | karpathy
  • supervised learning that is ~absent in the animal kingdom
    Animals vs Ghosts | karpathy
  • In Sutton's world view, all there is is an interaction with a world via reinforcement learning, where the reward functions are partially environment specific, but also intrinsically motivated, e.g. "fun", "curiosity", and related to the quality of the prediction in your world model
    Animals vs Ghosts | karpathy