Owner DWS: AS30179 | UIUC CompE | Current: @Welltower | Prev: engineering @datadog, @ https://nitter.gq/t.co/CGjfaWTie4
New York, NY
Joined October 2010
- Tweets2K
- Following430
- Followers208
- Likes2.7K
Seems like everyone is clamoring for cinematic grading with a slider.
How about just take the picture, no edit needed.
Amidol, a culmination of film simulation studies I’ve been doing for a few years, will be on the app store soon!
4
132
It seems like the only thing you need for this is a custom domain email, and half believable story.
My take is that Anthropic IPO numbers are bad, and it would look great if a whole bunch of "startups" (with minimum 2 seat 1 year commitment! (albeit free)) signed up right before their real/2026 S-1 numbers come out.
The Claude Startups program is expanding to more founders.
Members can get a year of Claude Team, API credits, special offers on the tools a startup runs on, and office hours with Anthropic’s Applied AI team.
2
93
Downloaded a tried some of these. Not reverse engineered, more like closed-box engineering with a 1000ft pole.
Definitely impressive that some folks to the effort to put this together, but it has all the common bugs of a vibe coded photo editor (most of which I was running into building a toy image viewer a couple years ago!)
For example, DNGs have a surface level implementation (grabs the first image page, instead of the actual RAW on open), self implemented lens corrections (has AI not heard of LensFun?), and I couldn't find a way to get a proper scene-referred workflow on a high-gamut display.
Again, props for starting up the project, and I expect I'll make a few PRs in the near future, but don't treat this as a drop in for the battle-tested tools just yet.
free and open source reverse engineered adobe products
getartcraft.com/apps
Readers added context they thought people might want to know
The linked projects describe their apps as clean-room reimplementations built from scratch in pure Rust from public specs and observed behavior, not reverse engineered.
github.com/storytold/phot…
github.com/storytold/vect…
github.com/storytold/vect…
github.com/storytold/phot…
getartcraft.com/apps
3
136
I'm pretty convinced that the future of DevOps/Infra is agentic. Too much work in the sector is just runbook based or "runbook discovery".
I think you can probably let a strong LLM go after a task, write a runbook, compress the runbook history to a minimum step DAG, and then let a weaker model run the DAG + diagnose errors.
I've started doing this in my homelab and my entire maintenance burden has effectively dropped to 0. The key, however, is that you must codify the runbook. A prompt is _not_ a runbook - it is a chance of generating the right runbook.
There is probably a broader lesson in LLM usage here that others have already figured out, but prompt -> steps -> runbook -> DAG -> DAG + Prompt seems to be a pipeline that is highly maintainable, observable, and predictable in results.
1
32
If it ends up being true that OpenAI used a specific user’s session to assist in their own attempt at solving Navier-Stokes, then the enterprise community should be very concerned.
The implication is that OpenAI is able to target a specific user, grab their data, and use it to their advantage.
Smells a lot of 2010s Facebook: “oh we have lots of privacy safeguards in place! we can’t look at your chats!”. Meanwhile….
>math professor spends a year on one of the hardest unsolved problems in math
>"Navier-Stokes"
>his drafts for the whole project went through codex sessions
>openai had his logs in codex
>suddenly rumor leaks AI solved it
>he emails openai to ask what's going on
>openai calls within days
>"funny story, our model also solved it"
>using his exact approach
>asks openai: did you train on my sessions?
>no answer
2
229
Tanishq Dubey retweeted
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!
The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.
125B parameters + 51B N-gram embeddings, with just 6B activated per token. Unmatched cost-efficiency.
What's new: 🥳
- Next architecture: GDN + QSA hybrid attention, Gated Residual, N-gram Embedding & Muon optimizer, serving as a precursor to the architecture used in Qwen4.
- Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks.
- Strong performance: scoring 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI).
- 262K native context, extensible to 1M with YaRN.
We’re also releasing the weights for Qwen3.8-Flash-Next, giving the community an early look at the new architecture we’re exploring for Qwen4.🚀
We can't wait to see what you build with Qwen3.8-Flash!👀👇
- Blog: qwen.ai/blog?id=qwen3.8-flas…
- Technical Report: github.com/QwenLM/Qwen3.8-Fl…
- Hugging Face: huggingface.co/Qwen/Qwen3.8-…
- ModelScope: modelscope.cn/models/Qwen/Qw…
> Working at Valve taught me that "cleanliness" is not the most important factor for success
Important lesson for engineers imo
That looks like vgui code.
vgui is the absolute worst UI system anybody ever wrote. Which of course, meant that we* had to build a bunch of games on top of it.
* Mostly before my time.
Working at Valve taught me that "cleanliness" is not the most important factor for success
1
141
Qwen3.8 27B seems to be a fantastic model checkpoint based on benchmarks. A nearly 20% jump in computer use means that this should be excellent for agentic applications!
We promised open weights for Qwen3.8. Now, time to meet them! 🎉
⚡ Qwen3.8-27B:
- A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.
- 262K native context, easily extendable to 1M tokens via YaRN.
- Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0.
🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently.
Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now!
Download, deploy, and build something we haven't imagined yet. 👀👇
- Hugging Face：
huggingface.co/collections/Q…
- ModelScope：
modelscope.cn/collections/Qw…
4
347
We're just a couple hours away from Qwen3.8 27B being released, so DWS has something special for the community:
A speculative-decoding engine that makes MLX Qwen3.6-27B run ~1.7× faster than mlx-lm's own greedy decode on Apple Silicon. The fastest Qwen3.6 inference on Mac!
And, this isn't a benchmaxxed result either. We maintain our performance gain through 128k context. More in the thread ->
And we measured this result at various output lengths as well, ensuring that the engine's performance is consistent.
1
1
169