Posts on X before Oct 23, 2024, 8:05:49 PM UTC

coming soon to obsidian (@obsdmd) x systemsculpt ai plugin :) the way that @kepano made canvas is so intuitive, easy to understand, and so comfortable to mod on. lots of work for me left to do before a full release but it's looking like sometime next week!
Media attached to this post
github copilot seems to be just as useless as it was back in 2022. the only useful feature it has is generating commit messages (for some reason @cursor_ai never answers why they don't implement their own little "Generate Commit" button). anyway, back to Cursor/Aider.
DOM-use, inspector-use, terminal-use, API-use all would be much better.
computer-use honestly doesn't seem to be a very smart direction. we probably need to build separate interfaces for llm agents.
if they released this new sonnet as opus 3.5, it’d be a little underwhelming BUT people would still pay the premium for the extra brain power it has over the original sonnet 3.5
holy fuck. the new sonnet figured something out that I've been trying to do for the entire day today... in 1 shot.
things sonnet 3.5 new has over o1-preview: - accepts images/docs - streaming - is a doer, not a thinker (no 20 second think times) - 200k window vs 128k - $3/$15 input/output costs vs... $15/$60... LMAO - new computer interaction paradigm it's anthropic's MID tier model. SotA.
Anthropic btfo’s openai o1-preview
Media attached to this post
these 3 things do become true if you just end up working for yourself / your tpot oomfies
some of ya'll "cracked" early 20's SWEs are going to be in for an awakening when you're forced to learn that: - TPOT != IRL - your interviewer isn't going to agree that "you can learn anything in 2 weeks" and - most of your co-workers aren't going to find the edgy comedy as funny
introducing our new MEGA-SWAG-V4 model, which extends DONKEY (based on GPT-GOKU4), with state-of-the-art D4NCE-RNG for stochastic dance synthesis
"our D̸̥̫̈́̇Ó̷̭̗̑̈́Ń̶̤̯K̷̢̈E̷̫̜͇̓̚͝Ý̶̩̠̽ method improves upon 𝐿𝑜𝓇𝒹𝐹𝒶𝓇𝓆𝓊𝒶𝒶𝒹 by replacing the SkibidiAttention blocks with a linear-time variant inspired by ░ᴋ░ᴠ░ ░ᴄ░ᴀ░ᴄ░ʜ░ᴇ░ ░ɪ░ɴ░ ░ʙ░ɪ░ᴏ░ꜱ░" x.com/jxmnop/status/…
nothing interesting happens on saturdays, sundays, or mondays. enjoy your 3 days, get off twitter, go build something, rest, spend time with your family
this is a beautiful representation of what most government agencies look like when it comes to efficiency
Along with our partners at @FEMA, we will continue with disaster recovery as a result of Hurricane Milton. The safety of the American people is our top priority.
you love to see it - this definitely increases the chance of us seeing opus 3.5 or something spicy from openai very soon
nvidia casually dropped an open 70B model that beats gpt-4o and claude 3.5 sonnet
Media attached to this post
oh look, another boring wednesday sitting around here waiting on my knees until thursday tomorrow for some opus 3.5. or maybe some sugar from openai. something's got to happen this week or I'm going to lose my marbles
i imagine that the windows pc experience for me feels like how the archlinux experience feels for windows users
I really hope this hot and spicy week of new frontier AI models starts today and not on Thursday or, God forbid, next week
every week... o1 eats and learns... every week, the list of banger AI/ML research papers increases. openai already has an MLE benchmark which attempts to apply ML learnings into the real world. it's also confirmed by many openai personnel that o1 is already self improving.
🚨This week’s top AI/ML research papers: - Differential Transformer - GSM-Symbolic - Pixtral 12B - Intelligence at the Edge of Chaos - Cheating Automatic LLM Benchmarks - nGPT - Upcycling Large Language Models into Mixture of Experts - Personalized Visual Instruction Tuning - Tow… MoreLess
🚨This week’s top AI/ML research papers: - Differential Transformer - GSM-Symbolic - Pixtral 12B - Intelligence at the Edge of Chaos - Cheating Automatic LLM Benchmarks - nGPT - Upcycling Large Language Models into Mixture of Experts - Personalized Visual Instruction Tuning - Towards World Simulator - Only-IF - Addition is All You Need for Energy-efficient Language Models - Selective Attention Improves Transformer - MLLM as Retriever - Rectified Diffusion - Everything Everywhere All at Once - Astute RAG - LLMs Are In-Context Reinforcement Learners - Scaling Laws For Diffusion Transformers - EVOLvE - Rewarding Progress - Falcon Mamba - Efficient Dictionary Learning with Switch Sparse Autoencoders - Scaling Up Your Kernels - RL, but don't do anything I wouldn't do - Aria: An Open Multimodal Native Mixture-of-Experts Model - Inheritune: Training Smaller Yet More Attentive Language Models overview for each + authors' explanations read this in thread mode for the best experience
Media attached to this post
agreed, anthropic has remained king for 4+ months while openai is doing backflips and cartwheels trying to impress everyone with "look, it talks! natively!", "look, it thinks! reasons in the background!" and yet nothing beats sonnet 3.5 as of today.
nice chart you got there. unfortunately, my technical analysis is superior x.com/EpochAIResearc…
Media attached to this post
if only apple allowed an easier way to use apple notes and apple reminders through a simple API... it would kill so many startups that are currently trying to patch this issue with third party software
I've said it before and I will say it again, light mode is literally slept on. Anytime I can turn on light mode for something, since a lot of apps default into dark mode nowadays, I do. I don't know what happened, but some time in the past few months, dark mode on things just doe… MoreLess
I've said it before and I will say it again, light mode is literally slept on. Anytime I can turn on light mode for something, since a lot of apps default into dark mode nowadays, I do. I don't know what happened, but some time in the past few months, dark mode on things just doesn't keep my attention as well as light mode does.
im really trying to come up with a mnemonic for uithub... maybe u for understand? Understand a GitHub. yeah, that works, thanks twitter void
can't wait for full o1 and opus 3.5 to hit before the holidays. with prices dropping, we'll hit the new year with so much more intelligence for pennies on the dollar vs. what we already have right now
When I browse the For You page, I can feel every time some of my brain cells commit harakiri because of an engagement rage post
I just want a think_time API setting for OpenAI where I can say “think for at least 5 minutes” or whatever, since they said the longer it thinks, the better the answer
At this point I need to stop using For You on X, clean up my follows list, and just stay in the Following tab All I see is engagement bait with 10K+ likes, all noise no signal
My reason was I was sick of Jagex fucking up RuneScape back in 2006, I found out about private servers / Java / bash, and the rest is history
i only started coding because i wanted to make gta san andreas mods btw
Welp, there go roughly 200,000 jobs (stat from US Bureau of Labor Statistics). Sure, you can argue that a lot of people go to chat it up/gossip/socialize in general at the nail salon, but I think we’re headed towards a generation where the majority of communication will be done … MoreLess
Welp, there go roughly 200,000 jobs (stat from US Bureau of Labor Statistics). Sure, you can argue that a lot of people go to chat it up/gossip/socialize in general at the nail salon, but I think we’re headed towards a generation where the majority of communication will be done digitally, not in person. People are becoming more and more anxious, shy… I know speaking for myself and my wife, we both would prefer to go to a robot instead 🤷‍♂️
so supposedly something cool is supposed to drop from GitHub Copilot (remember that? lmao) as well as from Google, less confident about the latter being released today but GitHub CEO himself said today they're releasing something new
$100 billion here, $150 billion there, and everyone laughed at @sama when he said $7 trillion
Accelerate: "Microsoft and BlackRock are part of a group of companies collaborating to pull together up to $100 billion to develop data centers for artificial intelligence and the energy infrastructure to power them." cnbc.com/2024/09/17/mic…
Even if the jump from Opus 3 -> Opus 3.5 is half as powerful as the jump we've experienced from Sonnet 3 -> Sonnet 3.5, it'll be SOTA by far, even further than o1. I'd pay $50/$100 at this point, IDC. @AnthropicAI just please release it already 🥲