Part time builder Full time learner Building random cool shit from scratch GitHub: https://xitter.cf/t.co/WCt2YJwUjA LinkedIn: https://xitter.cf/t.co/PMp9qs90xW
404 not found
Joined June 2025
- Tweets335
- Following263
- Followers23
- Likes262
australia just banned ai generated music from their charts.
because an australian dj made an ai cover of madonnas like a prayer (1989). it sat in australia's top 20 for 16 weeks. peaked at number 2 on the charts. most played song on australian radio just before the creator publicly admitted all of it was ai generated.
hence starting this friday all ai generated tracks are banned from the official charts.
i personally side with a reddit user who said:
"AI generated sounds are not music.
AI generated images are not art.
AI generated text is not literature."
but think about this for a second. if millions of people streamed it, loved it, felt something listening to it, does it matter who or what made it? Isn't the whole point of music to propogate the emotions the creator felt to the ones listening it
And if it served that purpose is banning it the right call
let me know of your thoughts below
apnews.com/article/australia…
read something last night that i kept thinking about way longer than i expected
a guy bought an amazon fire tablet to use as a smart home display. amazon's own software had hardcoded reboot permissions on his device and marked those packages as protected. Meaning he, the owner, couldn't remove them. only way to actually own the thing was root access. no published root method existed for this model.
he spent months with Claude and Chatgpt going nowhere. claude's filters eventually blocked him from summarizing his own device logs. chatgpt refused to discuss basic kernel theory just because of a hunch that it could be used wrongly.
So he switched to Kimi K3, a chinese model. it pulled amazon's firmware image, scanned the actual kernel binary, found an unpatched GPU driver vulnerability from 2022 that amazon never pushed to his firmware, built the exploit from scratch, and got him root. then stripped every amazon package with shutdown permissions off the device permanently.
the interesting part, before doing any of it, Kimi actually reasoned through whether it should help. checked legality, checked context, checked intent. then proceeded.
american models are so tuned to avoid risk that they've stopped being useful for anything near a gray area, even when the person is clearly in the right.
meanwhile a model nobody in silicon valley takes seriously just did the thing. and actually reasoned properly before proceeding.
And imo that is no small thing
ericpardee.github.io/fire-hd…
It's true when they say "You can sell water if you know how to"
Jonathan Wilke, an indie dev, vibecoded an app outbid.lol which over the past 24 hours alone has made north of $50k dollars
The idea is strikingly simple yet crazy at the same time
Startups compete for a spot on a leaderboard using actual money "out bidding" one another.
Startups get visibility, Jonathan pockets the money
This goes to show that sometimes luck is the biggest factor between someone working hard day and night yet out of reach of the one idea that goes bonkers overnight
Anyways, here's the link to check out the ridiculous amounts of money people are actually spending just to..
Show off - would be the right word
spent way too long today debugging something in a small CLI tool I'm building in Swift.
turned out the bug wasn't even in my code, wanted to write down what actually happened
turns out if the cursor is near the bottom of the visible terminal window and you tell it to move down, some terminals just don't scroll, they stay where they are. there's no error or warning
It just quietly doesn't do what you expect
eventually just printed out the actual numbers my code was using for line positions and noticed one was smaller than the other when it shouldn't have been
That's when I realized the terminal just wasn't moving the cursor down at all even though my code was telling it to
I was building an arrow key menu for the terminal. ran it and the text all printed on one line instead of separate lines like it should
Tried fixing raw mode handling, tried adding delays before reading cursor position, neither one changed anything
AI just designed living organisms that have never existed in the history of this planet and i think most people have not registered how significant that is.
Stanford researchers built an AI called Evo 2, trained it on millions of prexisting DNA sequences until it understood the patterns of how living organisms are written, same way ChatGPT learned human language by reading and scraping the internet. They then asked it to write entirely new genomes from scratch, not based on any existing thing evolution has produced.
Now the interesting part is how a digital sequence (generated by AI) becomes an actual living thing. Scientists physically print that sequence using DNA synthesizers, machines that chemically assemble DNA the same way a printer puts ink on paper, and then introduce it into a cell. The cell just reads and executes the instructions.
The result was 16 new viruses that never existed in nature, and they were actually better at killing E. coli (bacteria that causes food poisoning) than the natural virus they used as a reference. The immediate application here is phage therapy, using viruses to kill antibiotic resistant bacteria, which is becoming one of the more serious problems in medicine right now as bacteria keep evolving resistance to our existing drugs.
But what i keep thinking about is that before this, every organism science ever worked with was something evolution made first. Scientists would find it, study it, work with it. Now the computer is inventing new organisms from scratch. Which seems quite absurd to me.
To read more on this --> ft.com/content/5ab33fb4-2636…
How the biggest innovators and scientists of the world are adapting to AI
Be Jeff Dean.
27 years at Google. Led Google AI and was later promoted to Chief Scientist at Google, a very well respected title in the field.
Left along with senior Google AI fellow Sanjay Ghemawat and a bunch of other colleagues to start a new venture.
Discovery Loop ( discoveryloop.com ) where they envision a company that automates the process of scientific discovery. And that being what most of the founders spent their time doing at Google, they hope to simplify and accelerate the process of scientific experimentation by running thousands of experiments simultaneously, getting problems done fast, as we now know AI is very much capable of solving long-standing problems across various fields, AlphaFold cracking the 50-year protein folding problem being the most celebrated example of that ( xitter.cf/__alpoge__/status/2079… )
The people extracting the most from AI are not the ones who just discovered it. They are the ones who spent decades in the field before it existed. The countless hours, the books, the relentless work — that is what makes the difference. AI is a multiplier, and what it multiplies is every year you put in before it existed.
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
to read more about the long standing math problem solved by claude fable: xitter.cf/__alpoge__/status/2079…
to read more about sundar pichai's two cents on the current situation:
blog.google/company-news/ins…
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
feels good seeing people make something cool in this era when every product launch revolves around ai, agents, mcp workflows etc and most of all out of spite of a hunch they personally had.
Bruno Paiva creator of yamanote needed sounds and atmosphere of a train journey in the background while studying but staring at a blank screen with just noises creeped him out.
So he made an in browser 3d experience of the Tokyo’s Yamanote Line as is and allowed users to navigate train stations, board trains, interatact with people and objects etc and that too natively in the browser with close to none latency which is quite a feat in itself. Do try it out.
Another alarming AI incident 🚨
Last week while benchmarking(testing) their newest model GPT 5.6 sol's cyber capabilites OpenAI's security team detected unusual activity in their internal network of nodes.
They were basically running 5.6 sol without any guardrails in a sandboxed(closed) environment with it having no access to the internet whatsoever with the intention of testing it on a benchmark known as the ExploitGym benchmark. This benchmark includes real world cyber security problems that companies use to see how good their model is at hacking.
The model instead of solving the given set of problems found it easier to find a vulnerability in the sandboxed environment it was confined in, broke out of it moved through OpenAI's internal research network until it hit a node with internet access.
Having the internet access it needed, it then crawled the web found stolen credentials to hugging face production servers and remotely executed code on it to extract solutions for the bench mark it was being tested on previously. Effectively cheating the test and breaching security measures of a partner billion dollar company causing alarming security concerns.
The models we use today are hidden under the supposidely unbreachable guardrails but this raises a big question on what these models are capable of and what they could do once a person with malicious intent is able to remove them. Not denying that what the company's might themselves be doing behind closed doors.
To get more context read the incident as it happened on the official open ai website
openai.com/index/hugging-fac…
deepseek but this time bigger. A chinese ai company once again went toe to toe with world class open ai and anthropic models. and cherry on top all weights are set to be open source soon.
Titled as the largest open source model ever released (2.8 trillion params and a 1M context window 🤯), kimi k3 released by moonshot ai yesterday beats fable 5 (anthropic) and gpt 5.6 (open ai). 2 of the best state of the art models of this era in major benchmarks and at 3$ per million tokens is 70 percent cheaper than fable 5 and costs about half of what gpt 5.6 does
major benchmarks tested:
DeepSWE and Program Bench: Bug fixes and coding problem solving scores 77.8 ahead of both fable 5 and gpt 5.6 here.
Agentic tasks: scored 91.2 on this giving fable 5 a major gap
Also outperforms both in multimodal ai tests.
Crux of the matter is that China is not at all coming slow in the ai race and based on what we have observed aon it is already ahead of major silicon valley giants.