building https://teapawt.net/t.co/XkSWnnq6ra, private personal software, built with your own AI harness. dev who complicates coffee more than it needs to be.
Joined December 2014
- Tweets216
- Following434
- Followers30
- Likes240
this is the problem we solved with fused-render.
Now shipping that power as a bot that runs locally 🚀
dm me to have a pre-release version
I'd call this a bubble in a bubble
Tech people somehow still think normies are not spinning up personal software already
They're not doing it in Claude Code in a terminal like tech people for sure
But their AI chat app is already doing it for them without them even realizing they're building personal software, like generating PDF reports, full web app dashboards and even entire web apps
That will keep getting more advanced and soon ChatGPT will just spin up a Photoshop clone (but ofc only if they need it, preferring to just do their photo edits for them)
Everyone will mostly use personal ephemeral software inside AI chat apps I think
Some software moats will remain like social media (community), ecommerce (think like Shopify and Stripe), etc but they will just be used by AI to build the personal software with
So maybe that's the point here, become a tool agents use to solve their user's problems and charge agents for it (like many are saying)
this is very relevant for anyone working with local AI models
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50.
Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model.
If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
vasu retweeted
Samsung just dropped a paper showing how to run a 13B model in under 1GB using extreme quantization.
If this hits production, local AI changes forever.
To put that in perspective, this is like running Kimi-K3 on a single DGX Spark.
Samsung open-sourced a method that shrinks 13B parameter LLM into less than 1 GB.
It's called "LittleBit"
Instead of storing AI weights as standard numbers, they used latent factorization to crush them down to extreme sub-1-bit levels.
In some configurations, they hit 0.1 bits per weight.
But here is where the architecture gets crazy.
When you compress an AI this much, you can stop doing math.
Instead of forcing the hardware to do heavy floating-point multiplication, LittleBit replaces the core computation with a bitwise XOR operation.
It swaps complex matrix math for basic sign flips.
The results rewrite the rules of model deployment:
• Unlocks a massive 11.6x inference speedup relative to standard FP16 models.
• Radically reduces memory footprint and loading bandwidth.
• Maintains robustness in extreme sub-0.5 bit regimes where previous compression methods catastrophically fail.
This is not a clever optimization.
It is the blueprint for running massive, state-of-the-art AI locally on cheap, resource-constrained devices.
best approach to hiring, just putting things out there 🚀
We're hiring software engineers (and interns) for our product team.
We put up two new $500 bounties as hiring challenges:
> github.com/commaai/connect/i…
> github.com/commaai/connect/i…
Put up a PR that we can just hit merge on, and the job is yours.
Seems possible for smaller projects. Bigger projects still have the the platform drifting issue where your business logic or some edge case handling will keep getting dropped by the agent in sync process.
I think where react native is right now is the best of both worlds. You should checkout the new @expo modules. They allow you to have native code right alongside your js which is pretty cool.
vasu retweeted
Nobody talks about it... but a lot of builders are quietly burning out right now.
AI was supposed to make this easier.
In some ways it did. We ship in days what used to take months.
But here's what I keep hearing from founders and engineers:
➣ Shipping faster doesn't mean you have more customers
Everyone has the same AI tools, so everyone launches the same product.
➣ Speed is exhausting
You can build all day now. So you do. And you're tired by lunch.
➣ Every new model release feels like a threat
"Will my product even matter after the next model launch?"
➣ Distribution got harder not easier
More launches, more cold emails, more AI content. Less attention for everyone.
➣ The internet feels louder and less real
Half the timeline is people selling courses about building not actually building.
➣ "Build in public" turned into "perform in public"
Fake screenshots, fake revenue, fake overnight wins.
But here's what still works from what I've seen:
- Solving one painful problem for one specific group of people
- Talking to real users not just shipping features
- Taste. AI can copy features but it can't copy care
- Trust built slowly, posting real wins and real failures
- Taking breaks before you burn out not after
AI made building cheap.
It made being trusted expensive.
If you're feeling this, you're not behind. You're paying attention.
What's been the hardest part of building this year? 👇
We are building fused bot which can do this in less than a minute.
Agents create apps for themselves and you, so things just work. All local, privacy first.
Launching soon!
Gave the same prompt to ChatGPT Dot, Muse and Instinct: Can you download this video [link from Tiktok]
🛑 Dot took 4 min and told me its access is blocked.
✅ Instinct didn't say a thing and just downloaded it in less than a minute.
✅ Muse installed yt-dlp, asked for permissions and got it in about 1 min.
Goated
be Sigil Wen
- 17: Moved to SF, slept at WeWork.
- 17: Joined Midjourney’s original 20-person Discord.
- 17: Met Chris Dixon through Naval.
- 18: Lived in an AI hacker house with Karpathy.
- 18: Became one of Claude’s earliest testers.
- 18: Helped Naval build Airchat and hire engineers.
- 19: Joined Spearhead as its youngest member.
- 19: Ran GPT-2 locally on an Apple Watch.
- 20: Got O-1 and EB-1A approval.
- 20: Founded Extraordinary to help founders get O-1 visas.
- 21: Became a Thiel Fellow.
- 22: Founded Conway Research.
- 22: Launched Underdog, a personal AI running on Macs.
- 22: Raised Conway’s first round, led by a16z.
can we have ios widget like shortcuts that appear on homescreen and opens the wabi apps directly.
Wabi apps now appear as native apps that the agents author and agents can access
I love the new wave of optimised software of daily tools.
Democratisation of software 🚀
WhatsApp Web takes 1 GB of RAM.
ZapFast is WhatsApp, native and fast: about 200 MB, opens instantly. Linux, macOS, Windows. Free and open source.
Fast to ship too: I released 3 versions in 5 days.
zapfast.rocks
That’s what i have been working on, you can try fused-render.
The idea is to empower everyone to build hyper personalised apps that solve their problems. We take the tech and the distribution away, so anyone can build
Hosting a live stream of h3 conference in blr that we are organising in nyc along with fused.
luma.com/exvzjqpi
Register and let me know, i will approve.
