Jonathan

https://kiro.dev/blog/introducing-kiro/ Days without new AI IDE: 0 Interesting concept tho. The basic idea is to break every task down into smaller and smaller subtasks to implement. This might be a more scalable architecture, since you can assign agents to different tasks simultaneously. It also probably scales better with AI capabilities since the AI can break each task down into fewer larger subtasks as model capabilities increase.

https://reason.com/2025/07/14/in-just-1-year-134-lifeguards-cost-los-angeles-taxpayers-70-million/ Wow apparently I’m in the wrong line of work. I should have chosen lifeguard as a career and made over $500k+ annually as an LA lifeguard. Just nuts.

Is there an email client for Linux that just works with Microsoft Exchange email accounts easily? That’s the last thing keeping me from jumping ship from macOS.

https://aider.chat/docs/leaderboards/ Grok 4 benchmarks are out for coding. Decent model, but not ground breaking. Grok 4 is more expensive than Gemini 2.5 Pro and o3 while performing worse.

https://moonshotai.github.io/Kimi-K2/ Really interesting model release from the Chinese lab Moonshot AI. One of the first large open source models I’ve seen that has strong tool usage. For general usage, not sure how good it is, but for agentic workflows it might be SOTA. Be on the lookout, because while this model is solid, the really exciting one will be when they add the reasoning and image understanding to the model.

https://reason.com/volokh/2025/07/07/book-on-machine-learning-is-full-of-made-up-citations/ Why is it that when people get caught using AI they just use more AI to cover up for their last usage? Lawyers do this exact same thing when they get caught submitting AI briefs.

https://asia.nikkei.com/Business/Technology/Artificial-intelligence/Positive-review-only-Researchers-hide-AI-prompts-in-papers I wonder if we’ll start to see people fighting back against LLMs like this. Putting little tricks in text to trick AI and doing poison pills.

https://x.com/adamscochran/status/1942113153245942233 Given the recent blatant lies from the head of the FBI Kash Patel about Epstein's death, the doctored footage, and the discrepancies between the government report on the incident and the actual footage I see only two possibilities. 1. Kash Patel is so stupendously incompetent that he on Joe Rogan he lied that the cameras not recording was fake news even though it was in the FBI's report and we only have footage from the one camera actually recording where you can't even see Epsteins cell. Kash Patel being stupid is a possibility, considering he's basically an aide who got to the top by brown nosing Trump. 2. Donald Trump killed Epstein. Most of the other candidates (Gates, Clinton, etc.) don't make sense because if someone killed Epstein it was someone who Trump and Patel would want to cover for. Clinton and Gates are definitely not on this list. Trump was also the president when Epstein died, so he was the person who could have done it the easiest. All of this is just so weird. I was pretty convinced that Epstein killed himself when all the right wing people came out and said that he did, but with all of these weird gaps and blatant lies in the FBI's official story this just gets way more confusing.

https://github.com/BlueFalconHD/apple_generative_model_safety_decrypted Well this is interesting. This guy decrypted the safety filters Apple has for their AI models. Apparently Apple is really concerned with anything relating to suicide, controversial politicians, and correct capitalization of Apple products.

https://www.osmo.ai/ Smell for AI. Makes you wonder, what do you think are the most important and information dense senses? My guess would be: 1. Sight 2. Hearing 3. Touch 4. Smell 5. Taste

Why is Eliezer Yudkowsky considered an important figure in the AI space? He's written some fun fanfic and a few interesting blog posts, but he's never actually done anything. Yudkowsky just seems high on his own supply of how intelligent he is, even though he's never accomplished anything substantive in the field of AI alignment or advancement.

https://marginalrevolution.com/marginalrevolution/2025/07/genetic-counseling-is-under-hyped.html Genetic counseling makes a lot of sense. If you’re thinking about getting married, many couples do marriage counseling. Why wouldn’t you get genetic counseling to see if your genes are compatible?

https://archive.is/2025.07.01-015114/https://www.nytimes.com/2025/06/30/world/canada/calgary-fluoride-water-canada.html To the surprise of what should be no one, it turns out fluoridated water is an incredibly effective measure to prevent cavities and removing it harms children’s dental health. Fluoridated water is one of the best low cost public health measures to increase dental health with nearly no downside.

https://worksonmymachine.substack.com/p/mcp-an-accidentally-universal-plugin The blog post’s point about MCP servers being a universal protocol for doing stuff on other platforms cracks me up because we already invented this. It’s called a REST API. The internet collectively spent 15 years hooking everything up to APIs before companies realized that they weren’t making money from their free APIs and shut them down (think Reddit, Twitter, etc). Now all the LLM agent hype is just making everyone remember how awesome it is to be able to do stuff on a platform without being forced to use some proprietary app. At least for now, it seems like we’re seeing a resurgence of APIs and more open and accessible data which I’m all for as long as the hype lasts before companies go back to locking everything down again. https://youtu.be/BxV14h0kFs0

https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/ The Gemma 3n model has so many optimizations that it’s hard to keep track of them. Take a read through the post and you’ll be up to speed with nearly every memory and compute optimization that’s been invented in the last few years. I think the most interesting feature is the MatFormer architecture. The cool thing this lets you do is that just like with Matryoshka embedding where you can lop off the last part of the embedding and vary the size of the embedding depending on how accurate you need it to be, the new architecture lets you vary the size of the model on the fly depending on how much memory and compute you have available. AFAIK this is a novel architecture for LLM models. Also side note, kind of embarrassing that after all the stupid games that Meta played to get their huge crappy Llama 4 models to the top of the LMArena Gemma 3n still beat them with a tiny model.

It’s weird to me that the most popular and intuitive programming language in the world—Excel—is data first and logic second, and then all the programming language designers just went “huh” and kept designing logic first data second languages.

https://github.com/google-gemini/gemini-cli Google is releasing ANOTHER tool with a free tier. The insane amount of money they are burning letting everyone use their models for absolutely free is nuts.

https://www.pewresearch.org/short-reads/2025/06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/ Good reality check, it may feel like LLMs are the whole world, we’ve just barely passed a majority of people ever having used ChatGPT, let alone using it routinely.

https://www.cremieux.xyz/p/columbia-is-still-discriminating Another admissions data hack has hit an Ivy League school and to literally no one’s surprise Columbia is still discriminating on the basis of race.

I've been living for two days in the overlapping area in a Venn diagram of the absolute worst kinds of debugging: network firewall config and Python conda dependency hell. With all the new Apple Health mental health stuff they should add a feature that's like "It looks like this is the third time you've tried to prune the conda base environment, and you're starting to swear in your messages to the LLM. I'm phoning the suicide hotline now."