I used ai to convert my resumé to html and have it as an endpoint on my GitHub.io page So I can use GitHub copilot to spin up a new variant very easily from the web And the html is structured so it generates a nice pdf using the browser print function

Microsoft CEO Satya Nadella says people should move beyond calling AI 'slop' "We need to get beyond the arguments of slop vs sophistication and develop a new equilibrium ... that accounts for humans being equipped with these new cognitive amplifier tools as we relate to each other"

Erroneously thinking you've taken intelligence tests, when the tests were administered because doctors are (repeatedly) concerned you have dementia and may pose a danger to yourself and others, is probably yet another sign that you actually do have dementia.

best way to end day 2: chicken roast and aloo bhortha from boishakhi

When a chatbot says it’s sorry it’s improv When a chatbot says it’s sorry it’s improv You can never change or guilt it Just go find the guys who built it When a chatbot says it’s sorry it’s improv

Happy to share a project I worked on finally. I found that you can cause a base model to behave like a chat tuned model, including using proper stopping tokens, using nothing but a series of vectors applied within the model's layers. The vectors are trained with descent on a chat dataset, like SFT.

Here is the editor of Tablet agreeably reposting someone calling nonwhite NYC immigrants "worthless biotrash." It's astonishing how quickly this stuff has escaped containment

Reminder to BlueSky newcomers—if you want to immediately fill your feed with NFL talk ahead of the playoffs, subscribe to my starter pack! Adding new accounts all the time 👇 bsky.app/profile/did:...

It’s funny wearable ai devices will be everywhere soon, but Humani ai pin was a massive failure They had the right trend but their timing was just bad It’s interesting watching people be comfortable getting recorded 24/7 too… 9to5mac.com/2026/01/02/a...

I've been pretty bullish on world models for a while

BREAKING: OpenAI is predicted to buy Pinterest, $PINS, in 2026, per the Information

1) this article is hard paywalled, so there's no way most of the people commenting on it have read it, including me 2) i've seen how people come up with baby names and seen what those names end up being, this is not particularly egregious in either case

This is pretty nuts. Someone used AI to write an article on fireworks that hallucinated a show at Brooklyn Bridge: www.timeout.com/newyork/news... and thousands of people showed up and got no show: www.reddit.com/r/TrendingAn...

Pretty much sums up where the USA is today. If Trump and his billionaire media tycoon buddies get their way, in 2026 we won't even be able to watch legitimate news from overseas. They are pushing for all alt-right propaganda all the time.

this is a very important paper, although not easy to read it takes hyperconnections (HC), which is basically just a smarter way to do back propagation, and stabilizes it so that it’s actually usable at scale

We don’t need “good immigrants “ etc stories We need a place where people get to be people and we don’t become cops and regulators monitoring everybody’s everything for signs or authenticity

Like most human beings, chatbots struggle to make new discoveries. Fortunately there’s a tiny irritating minority of humans who do make discoveries, in part by systematically doubting received explanations. Not clear whether it would be possible or ethical to increase their number.

The best open-weight coding LLM? Another Chinese quant company, like High-Flyer/DeepSeek, decided to enter AI and released the IQuest-Coder-V1, which is the best coding open-weight LLM per them. Blog: iquestlab.github.io Models: huggingface.co/collections/...

The Korean government told major tech companies to develop AI models ( www.koreatimes.co.kr/business/tec... ) and one of the five team released a new model. NC AI's VAETKI: A MoE model that 112.2B in total and 10.1B activated Model: huggingface.co/NC-AI-consor...

Happy Gary Marcus was wrong about 2024 and also about 2025 day for those who celebrate.

Based on data from 27 reports, basic income recipients are more likely to find long-term employment, and not a single pilot has resulted in reduced work. Recipients also experience less stress, because of course they do.

Public views on AI are much more muddled and less ideologically polarized than you'd think from browsing bluesky

Wow! This is quite a jump! More so than what I expected.

BREAKING: The House Judiciary Committee has released Jack Smith's 255-page deposition transcript: judiciary.house.gov/sites/evo-su...

Biggest flaw on iPhone 17… too many buttons. No mute switch

SNAP already doesn't allow people to buy hot prepared foods. It's a paternalistic program and because it's targeted by income, it creates divisive stigma and also imposes 30% marginal tax rates on recipients. The way to improve it is do cash instead, remove conditions, and universalize it, thus UBI.

BREAKING NEWS: Actual facts appear for the first time on "Truth" Social.

China has drafted strict rules to stop AI chatbots from manipulating vulnerable users regarding self-harm. If suicide is mentioned, a human must intervene immediately. Minors and elderly users must list a guardian to notify if self-harm comes up. I hope this becomes the global standard.

I think this is exactly the math that is going to drastically change the small SaaS landscape why would I pay monthly for <tool> when the same cost (maybe less!) can make that and also anything else I need

+1 just for the mermaid diagram

Big redistribution of wealth from the first world to the third world going on rn

Survey on vision encoders in VLMs They looked at 70+ models and found that training methodology beats scale. A well-trained 400M encoder outperforms a 6B one. Native resolution matters for docs, multi-encoder fusion works.

This is kinda how I feel about ai coding The ai wouldn’t have ever made the thing I asked, but I never would have taken the initiative to build 90% of the things I make with ai

It's nasal!

Tencent released Tencent-HY-MT1.5: Translation models that translate across 33 languages and 5 Chinese dialects. - 1.8B On-Device Power: 1GB memory footprint. Using on-policy distillation to align with larger models, it delivers 0.18s latency (50 tokens), outperforming mainstream commercial APIs.

Recently, LLMs were found to encode different languages similarly, a sort of Platonic representation of words. It now extends to science:: 60 ML models for molecules, materials & proteins (all with different training) converge toward similar encoding of molecular structure arxiv.org/pdf/2512.03750

Someone has lifted the QuickTimeVR movie files from 1994's "Star Trek: The Next Generation Interactive Technical Manual", upscaled them then built an in-browser replica of the ITM's "Tour" module. mijofr.github.io/st-panorama/ #StarTrek

A prediction I’ve seen a lot of recently is people will turn back to institutions because of AI— relying on trusted experts to determine what’s real and what’s fake Same principle could apply to other systems, relying on in-person and analog techniques

The world’s largest accounting body will end online exams due to rampant cheating using AI tools. I think this will be the endgame for any entity that wants to preserve the integrity of its testing processes including job interviews. It’s too difficult to outpace AI-based cheating at this point.

Turbo model of flux2dev that should be able to run on a lot of Macs Might be time to resurrect some of my old stable diffusion servers 🤔 huggingface.co/fal/FLUX.2-d...

A vibe-coded google earth flight simulator by William Holmberg. github.com/WilliamAvHol...

I’ve never been one to keep an ideas notebook but I love that ai has become my defacto idea notebook Only problem I have now is ideas are scattered throughout different providers— I need like an app to unify all my ai accounts…

An actually unique take on AI from one of my high school friends and his friends…

Meta acquired Manus

Gang please don't use ChatGPT to write your acquisition announcement manus.im/blog/manus-j...

this guy has his own heavily neuro-inspired model arch, NeuroBlast, that he’s designing at tiny scales using the SYNTH dataset from Baguettotron read more here: open.substack.com/pub/mkurman/...

For actual hard programming problems I’m getting much better results with codex 5.2 than either opus or Gemini high models

The “opposite” of a GPU is kinda weird.
 Graphcore has a strange chip (IPU) that loves tiny batches...but is also massively parallel. It used to cost ~$100k. Now you can get one on Ebay for $150 bucks. The catch is it's almost impossible to use.

I went to Texas to write about the sudden end of one of America's most successful experiments: the six decades of welcoming India's most educated and ambitious citizens. www.nytimes.com/2025/12/29/o...