In January 2025, I had a fun hypothesis for a blog post : can LLMs write better code if you keep asking them to “write better code”? That was prior to the advent of robust agentic coding, but Claude Sonnet 3.5 was still able to iteratively improve on algorithmic Python code. The “better” instruction turned out to be underspecified: Sonnet abused that ambiguity to instead add a ton of useless featu

Cognitohazards can affect machines too. On Sunday night, Anthropic researcher Levent Alpöge casually tweeted a surprisingly simple counterargument to the Jacobian Conjecture , a mathematics problem that has been unproven for over 80 years. Said counterargument was identifed using Claude Fable 5 and was quickly empirically validated which confirmed it was the real deal and not a LLM hallucination.

Subscription-based coding agents such as Claude Code and Codex famously have 5-hour and weekly quotas on their LLM usage. Both of these are understandable: 5-hour quotas help stagger usage so the servers don’t get overloaded, and weekly resets prevent users from dumping an entire month’s worth of usage into a single day which a) also prevents overload and b) stops the user from just unsubscribing

Why Hy3? OpenRouter is a service that provides access to most LLMs with a singular API, which has become exceedingly useful as of late given the rapid cadence of new LLM releases. Due to the company’s role as an intermediary between users and the LLM APIs, OpenRouter has robust, representative data on how users interact with LLMs and it publishes this data on the AI Model Rankings page: a welcome

No vagueposting here, just look at the Estimated Read Time. You’ve likely seen many blog posts about AI agent coding/ vibecoding where the author talks about all the wonderful things agents can now do supported by vague anecdata, how agents will lead to the atrophy of programming skills, how agents impugn the sovereignty of the human soul, etc etc. This is NOT one of those posts. You’ve been warne

A month ago, I posted a very thorough analysis on Nano Banana , Google’s then-latest AI image generation model, and how it can be prompt engineered to generate high quality and extremely nuanced images that most other image generations models can’t achieve, including ChatGPT at the time. For example, you can give Nano Banana a prompt with a comical amount of constraints: Create an image featuring

Nano Banana allows 32,768 input tokens and I’m going to try to use them all dammit. You may not have heard about new AI image generation models as much lately, but that doesn’t mean that innovation in the field has stagnated: it’s quite the opposite. FLUX.1-dev immediately overshadowed the famous Stable Diffusion line of image generation models, while leading AI labs have released models such as S

Whenever a new large language model is released, one of my initial tests is to try and jailbreak it just to see how well the model handles adversarial attacks. Jailbreaking an LLM involves a form of adversarial prompt engineering to attempt to bypass its safeguards against prohibited user input such as prompts requesting sexual or illegal content. While most of the LLMs from top labs such as OpenA

It’s an adversarial question for LLMs, but it’s not unfair. Last week, OpenAI announced and released GPT-5 , and the common consensus both inside the AI community and outside is that the new LLM did not live up to the hype. Bluesky — whose community is skeptical at-best of generative AI in all its forms — began putting the model through its paces: Michael Paulauski asked GPT-5 through the ChatGPT

ChatGPT and Claude won’t, but Gemini will. I’ve been working on a pipeline for representing an image as semantic structured data using multimodal LLMs for better image categorization, tagging, and searching. During my research, I started with something simple by taking an image and having a LLM describe who is in it: if they’re famous, there should be more than enough annotated images in the LLM’s