Audio Machine Learning Engineer tinkering with backend systems.
Joined May 2024
- Tweets3.3K
- Following577
- Followers725
- Likes1.2K
Pinned Tweet
Created the Mario theme melody using nothing but sine waves and a frequency equation.
Here's how i did it.
• Took the Mario melody’s notes and rhythm.
• Converted each note into a frequency: f(n) = 440 × 2^((n−69)/12).
• Added sine waves at odd harmonics to create the chiptune tone.
• Gave each note a short fade-in and fade-out to reduce clicks.
• Scheduled the notes and rests to recreate the melody.
• Generated 48,000 audio samples per second ie. the audio was sampled at fs = 48KHz.
• Slowed the rhythm without changing pitch just for fun.
The equations synthesized the sound.
and viola! we have Super Mario bros. theme music
I prompted GPT-6.1 Sol, GPT-6 Astra and GPT-6.1 Terra to build a 3D rocket launch and landing in the browser.
Model Comparison:
All three used High reasoning. All three reached touchdown.
• Sol: 11m 10s · $1.32
• Astra: 12m 05s · $5.83
• Terra: 9m 30s · $0.28
Terra finished fastest and cheapest, but delivered simpler visuals and less extensive testing.
Costs are API equivalent estimates as i used codex and not my Codex bill.
Top: Sol. Middle: Astra. Bottom: Terra.
Which one actually looks worth the extra time and money?
If you're in tech, let's connect! 🤝
I'm mostly interested in:
🤖 AI & Machine Learning
🎙️ Audio AI & Speech Tech
💻 Software Engineering & Backend Systems
🧠 DSA & Competitive Programming
🚀 Building projects & sharing what I learn
By day, I work as an Audio ML Engineer at a startup.
Outside work, I love building side projects, exploring new technologies, and sharing my learning journey.
Let's connect with like minded builders, engineers, and curious minds! 🚀
Since this post a lot has changed.
I have made a lot of friends in a very less time, people highly active in
AI Engineering
learning and developing harnesses and its engineering, Devops and systems side of things.
Machine Learning
Audio Engineering
Inference Engineering
Lot of cool spaces to be in right now.
Would want to connect with loads of more people.
Haha , enough yap lets larp together :)
Gained 60 amazing connections today and yesterday,
I Love the fact connecting with people working in the same niche in tech.
Also the thing is friends I have made in just a month living far and wide away but still are a part of journey.
Looking to make more genuine connections and helping each other out in our journey to greatness.
Let's settle Kimi K3.1 Pro vs Opus 5.5 vs GPT-6.1 Sol.
Which one actually earns its price?
I checked published quality scores, benchmark costs and a real coding test.
Here's the compact breakdown, including the tradeoffs. 1/5
4/5
A real coding example, from Atomic Agent:
One Python task, 32 hidden tests, three runs each. All passed.
Kimi low: 132s, $0.16–0.21
Kimi max: 294s, $0.35–0.46
Opus: 69s, $0.37
Lower token prices didn’t guarantee a cheaper run. Sol wasn’t included.
atomicagent.io/blog/kimi-k3-…
My reading of these results/Conclusion:
Opus leads the AA aggregate score.
Kimi narrows the gap; weights are promised by November 4.
Sol has the lowest AA benchmark cost here.
Measure pass rate, total bill and time on YOUR tasks before switching.
What should we test next?
Source for weight availability:kimi.com/en/blog/kimi-k3-1-p…
Hedwig's Theme from Harry Potter.
Created using Sine wave and notes being played at diff frequencies.
Motion graphs using 6.1 sol High , used mere 10 percent of my 5hr quota.
Latest Codex Update: harness knows the amount of tokens left and keeps a track of it throughout the conversation.
Also tells you out Loud.
below image exclaims codex harness thinking out loud that it has 29K tokens left.
Mythos was hyped to end all vulnerabilities, but instead, it just dropped 20,000+ CVEs on us that nobody has the human capacity to patch 😭
And don't worry, if Claude faces a real security threat, we definitely won't find out from a leaked third-party vendor report weeks later... right? 🤡
