20 yo | founder @ ofcompute
- Tweets 2,225
- Following 584
- Followers 738
- Likes 2,191
we already saw this exact situation with @useKled and new with @UseFastlane
crazy how the copycats are not just copying the idea but their ips and designs too 1:1, like dude wtf
Google just led a $20m round into a complete rip-off of my startup, Fastlane.
(skip to 55 seconds to skip context)
This isn’t something I’d normally talk about, but a line has been crossed. This disgusting and egregious behavior honestly makes me sick.
Fortunately, there is nothing anyone could do to stop us from winning. Especially not some unimaginative NPC’s. Good luck fellas.
Community note
The post claims Google led a $20M round into a real "Relay" rip-off of Fastlane, but no such funding or scaled competitor exists; it amplifies unverified hype for a basic tool contradicted by its site and community notes.
relaystudio.ai
x.com/klashsarma/sta…
doomers.ai/launches/relay
3
345
We're in the third big shift in how software gets built.
1. Copilot in VS Code (2022-2025)
AI writes the code. You read every line and catch the bugs.
2. Claude Code (2025-2026)
You stop reading code. Prompt, check the result, prompt again till it's right.
3. Grok Bot (2026-now)
You stop watching. Each Bot gets its own cloud computer with a browser and a terminal. It logs into your tools, grinds through the task while you're gone, and only pings you when it needs a yes.
3
228
Introducing BotRoster an open source alternative to Dots by openai
(also to grokbot and muse)
Github: github.com/mandarwagh9/botro…
Available across Mac, Windows and Linux (mobile app coming soon..)
1
1
3
393
as absolute troll of a person @elonmusk is
bro went and got dot.com, as @OpenAI launched "dots" as a direct competitor to @grok
and dot.com redirects to x.ai/bot
Dots are here!
A new way to use AI that works 24/7 for you; get more of your time and attention back to work at a higher level.
openai.com/index/introducing…
2
7
579
after using @bot for some time now, i think its so underrated
why is no one taking about it ?
1
65
Announcing Cosign.
Silicon Valley knows people by watching other people know them. One person's visible belief becomes a signal to someone else, and the network updates its priors.
Ascendancy is a game of two moves: self-made wins and high-status vouches. The earlier someone is in their trajectory, the less proof there is to go on, and the more meaningful the vouch. Belief compounds.
People used to earn status through rites of passage like graduating college and landing prestigious jobs. People collect status in bits now, 24/7, often outside any formal system. Showing your work in public has become the new warm intro.
Vouches take many forms. A tweet, an invite, an introduction, an angel check. Being undeniable and getting credible people to say so is how you become somebody in Silicon Valley.
The game only works if new people can win. Everyone benefits when good people get pulled into the game faster.
If you believe in someone early, say so.
cosign.co
194
Finally I have my name as a username somewhere 😋
cosign.co/mandar
92
the ceiling nobody states: if two annotators agree 95 percent of the time, a model scored against one of them cannot measure above about 95 percent. past that point you are fitting one person's judgement calls, so measure agreement before you chase the last few points.
1
79
low variance is the number to read carefully: a judge that returns the same wrong score every time has zero variance. consistency and correctness are different properties. the cost line implies about 970 calls, so the follow up worth having is agreement with human labels.
We tested Jev against LLM judges on accuracy, repeatability, latency, and cost to see whether System One models could offer a new approach to agent evaluation.
179
the number that would settle this is time between interventions. by the rule of three, an uncut 4 hour run with zero interventions puts the true mean above about 1.3 hours at 95 percent confidence. real progress, and a long way from a day of unattended work.
1
99
the question this raises is how many human labels. agreeing on 45 of 50 sounds decisive but carries a 95% interval of roughly 80 to 97 percent, so a 90 percent judge and an 82 percent one look identical at that size. logging the score makes it checkable later.
Can you use Jev for Evals? Yes! Remember that a LLM Judge is also classifier*.
Make sure to test your classifiers against human labels and don't overfit.
Hope this helps!
* hamel.dev/blog/posts/evals-f…
71
Mandar Wagh retweeted
Tiny actuators.
Accurate down to about 25 nanometers.
Xeryon builds piezo actuators designed for positioning tasks where normal motors simply are not precise enough.
• Repeatability around ±25 nm
• Linear, rotary, and multi-DoF configurations
• Used in metrology, semiconductor tooling, and laser systems
At this scale, motion is no longer about speed or power.
It is about staying exactly where you are supposed to be.
——
Weekly robotics and AI insights.
Subscribe free: 22astronauts.com
19
118
1,485
326,989
Mandar Wagh retweeted
leaving a few beginner friendly guides for classifiers (normal and zero-shot) as well as where you can find them on the Hub
opt for DeBERTa and ModernBERT ones
> huggingface.co/models?pipeli…
> huggingface.co/tasks/zero-sh…
> multimodal (image <> text) huggingface.co/docs/transfor…
people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows
joke aside I always found zero shot classifiers to be fascinating and was sad we never got people to scale them as much as decoder only models
many problems solved with LLMs could have been solved with them, it was a skill issue
9
44
385
43,484
ironing a matte filament is self defeating and it is worth knowing why. the matte finish comes from a filler that scatters light, so re melting the top skin destroys the texture you bought the filament for, unevenly, which reads as banding. ironing is a PLA trick.
1
86