Building local AI agents and factory automation in Vietnam. 12 years shipping internal tools. Real tests, failures, and lessons.
Viet Nam
Joined September 2010
- Tweets1.1K
- Following384
- Followers94
- Likes3.8K
SteveB retweeted
We achieved SOTA on Terminal Bench 2.1 using a DeepSeek v4.1 Flash model.
92.2% at $0.08 cost per task.
Previously Opus 5.5 was the leader at 87.6% at $0.50 cost per task (vals.ai/benchmarks/terminal-…).
Now the obvious question is how did we do it? How can you possibly improve a Flash model to beat frontier models like Opus, Fable, and Astra?
The truth is, we over-obsess about models when they are only half of the equation. The other half is the code that wraps around the model. In fact all of the most revolutionary products (Instinct, Muse, Claude Code, OpenClaw, etc.) are all just harnesses optimizations.
Rather than try to build a new model or fine-tune, we built a harness that would make:
Flash Models —> Frontier Results
Frontier Models —> Beyond Frontier Results
Did I mention it’s also ~6x faster than Claude Code?
And we’re now live. Go Check it out!
Trymidas.io
SteveB retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/con…
SteveB retweeted
Introducing Jev @typesafeai powered social-api
Try it at treg.to/social
- One API to access any social data
- Track every mention of you & your competitors
- Find trending posts in your niche, filtered by Jev
- Under $0.05 to pull 1,000 competitor ads
Open source, 0% markup
👇 Repo link below
SteveB retweeted
Today we release Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass.
It can do all that in 74 ms on Samsung Galaxy S26 NPU or 107 ms on its CPU.
Mirage has a 7.8M encoder + a 40.8M decoder at 2bits, and describes images in English, German, French, Spanish, Italian, Dutch and Polish.
We trade general language capacity, to compete top small VLMs today (SmolVLM2-500M, LFM2.5-450M and Florence2-230m) on these tasks.
On 256 VQAv2 questions, feeding Mirage's structured output and full trace to text-only Qwen3-0.6B raised its score from 9.3 to 31.3 out of 100, without fine-tuning either model.
Read more and benchmark it yourself here:
cactuscompute.com/blog/mirag…
SteveB retweeted
an absolutely beautiful and informative blog post! one of the biggest anti-patterns we see with jev is just plugging it in to make Big Decisions instead of having code do what it's best at and jev for just the parts that require judgment 🤌
THIS is how we transform software!
SteveB retweeted
Voice direction & distance-aware UI.
The demo shows Voice Ring + Libraries Voice Position model, where the UI reacts to voice direction and distance. The ring also changes its color based on speech emotion. Ready to use with SwiftUI on Libraries.dev
SteveB retweeted
we made an open-source Grok Bot with unlimited usage!
meet @mausbot_
use your existing subscriptions: Claude, Codex, or even local models.
Try it here - mausbot.com
SteveB retweeted
LibreYOLO 1.7.0 is out, and this one is special: image to 3D!
The first model is Speridlabs Idol by @speridlabs, and the quality is stunning. One photo in, a 3D Gaussian splat out, in three lines.
Apache-2.0, weights straight from their Hugging Face.
libreyolo.com/docs/models/sp…
SteveB retweeted
Great to see Ming-3.0-tts-0.5B running locally on an RTX 3090! In @ItsmeAjayKV’s test, a 5-second sample of his own voice was enough for cloning—even with fan noise in the recording. 🫶 Thanks for sharing, AJ! We’d love to see what you build next.
Tested voice cloning with Ming-3.0-tts-0.5B by @AntLingAGI and it is even more insane.
For a model of this size, voice cloning happens really well, i tested my own voice, half drunk, in a room full of noise (fan noise mostly) and it instantly cloned it.
It needed just a 5-second sample of my voice, and it speaks faster than real time: about 4 seconds of speech in 2–3 seconds.
All running locally on my RTX 3090. Ming takes about 3.3 GB of VRAM, and the whole model is under 3 GB on disk.
Great time for Local AI.
SteveB retweeted
Your Harness. You Define It.
Meet Creator Mode in DeepSeek Harness.
Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer.
What will you create?
SteveB retweeted
Introducing Granted!
I bookmark a lot on X and almost never go back to it. So I built something that does.
→ Every night, agents read my X bookmarks
→ Research ideas, pick the best one
→ And build a working prototype.
Every morning I wake up to a startup brief + a live app on my phone.
Open Source, sharing soon!
SteveB retweeted
🎉 NVIDIA Nemotron 3.5 Transcribe is here—with day-0 serving through vLLM + NeMo Speech.
40 languages. Recordings up to an hour. Promptable transcripts, speaker labels and audio Q&A.
Open weights in BF16, FP8 and NVFP4.
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
SteveB retweeted
So one of our ambassadors @divagr1925 has been working on something cool
He found a way to use Qwen's Gated DeltaNet state as memory, and built a 32-bit RISC-V CPU directly into the model's weights. So instead of relying on predicting the answer to a calculation, the model can execute instructions internally. It can also run C and MicroPython programs.
He calls it Qwenputer, love the name btw. The results so far are pretty impressive. On one arithmetic benchmark, accuracy jumped from 48.6% to 100%.
Pretty cool to see developers experimenting with Qwen's architecture. Great work, my friend!
