I share interesting things I find, what I’m learning, and projects I’m trying. Follow me if you like tech, creative stuff, and learning new things. INFJ-T
Sinapore ⇄ Philippines
Joined March 2018
- Tweets1.2K
- Following1.1K
- Followers222
- Likes1.1K
Pinned Tweet
just released Last Manual Programmer.
A song for devs who remember writing everything by hand, now shipping with AI... but still reading the diff. 😂
Used Claude Opus 5.5 , Suno V6, Codex, ChatGPT
Aivan Monceller retweeted
FLUX 3 Image Klein is here.
Open Weights with the quality of a large frontier model. Customizable and fast.
5.6B parameters and runs on a single consumer GPU.
Native 4K, bounding boxes, and pixel-perfect editing.
AI changed how I choose a tech stack.
I used to pick what I knew, or what I was willing to spend time learning.
Now I care more about performance, maturity, ecosystem, tooling, and traction.
Less , what can I build with?
More, what's actually best for this?
As a Codex fan I haven't used any GPT models for anything serious since Opus 5.5 dropped. I now have a max 100 sub.
Been hammering it with projects in.. compiler, mathematics, gpu , porting, rewrites, multimedia, gamedev, systems prog, UI, tooling
anything I can think of
Aivan Monceller retweeted
Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
I like that opus 5.5 when tasked something very difficult.. it will not reach out for readily available libraries, it will try to build it on its own , it will prioritize no deps over having dep hell , it doesn't mean that it will not considering pulling a dep though.. I like this behavior
Even before Jev was released, LiquidAI released LFM 2.5 Encoder last July 2026 , and it runs on CPU , dn't sleep on @liquidai
Don't forget to claim your $250 docker sandbox cloud credits docker.com/c/sbx-promo/
Aivan Monceller retweeted
A new fine-tuning framework, Halo, just dropped!
And with it, two new recipes for fine-tuning our MoEs:
• LFM2.5-8B-A1B
• LFM2-24B-A2B
LFM2.5-8B-A1B: github.com/Liquid4All/cookbo…
LFM2-24B-A2B: github.com/whitecircle/halo/…
Introducing Halo, the best framework for post-training of open-source models.
Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format.
Star us on GitHub: github.com/whitecircle/halo
Step 5 is currently on preview , looking good on paper
600B params, 27B active, 1M context, vision, great at long running tasks.
Weights drop Oct 15 too definitely keeping an eye on this one.
Introducing Step 5 Preview: Advancing the Pareto Frontier.
Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance.
- 600B total / 27B active MoE, with 1M context + Vision
- Substantially lower task cost at comparable intelligence
- Broad software engineering capabilities with sustained execution over long horizons
Try Step 5 Preview: platform.stepfun.ai
Model page: stepfun.com/step-5-preview
Open weights on Oct 15.
Jev is free for a week, Build stuff and go crazy!
Jev by @typesafeai is free on Vercel AI Gateway until Sept 25.
Build with the fastest adopted model on the Gateway at no cost.
vercel.com/ai-gateway/models…
xitter.cf/vercel/status/21010773…
My custom harness is shaping up, off loaded most/all decision to @typesafeai Jev, let these local models focus on work, consult frontier when needed. It's a complete 360 , suddenly my harness became so smart overnight.
Aivan Monceller retweeted
Jev is such an obvious and useful “type” of model. It’s genuinely shocking we were this far into the AI revolution before someone shipped a general purpose classifier.
I bet @CompleteSkeptic woke up every morning and scrolled X half expecting to be ape’d. Well done @typesafeai.
Muse Spark 1.3 Contributor is kinda crazy cheap, 3 times cheaper output than DeepSeek V4.1 Flash off-peak. to be honest I feel DS4.1 Flash is still faster though. AAI Numbers tell they are just as fast and is more intelligent
this is an interesting shift in AI. The model is called Jev. What I find interesting with Jev is, right now most of our code is still deterministic or “frozen”, then AI is usually sitting outside as an assistant or agent.
if AI decisions become cheap and fast enough, you can start putting them inside the actual decision tree of the software itself.
So instead of having too many heuristics, thresholds, if/else, switches, etc, some of those fuzzy decisions can just be handled by the model.
hard rules can still stay as normal code of course, but the parts that are hard to hand code can become AI decisions.
So AI is not just helping the code anymore, it becomes part of the runtime logic itself.
You can read/learn more about it here
typesafe.ai/blog/introducing…
youtube.com/watch?v=cJ0EOzey…
