Engineering @TheTradeDesk / @Google / @Qualcomm ❤️Tinkering with 3D Printing / Electronics / HW / Rust / AI https://xitter.twitterapi.xyz/t.co/WaPyIkt8ay
Joined April 2008
- Tweets50K
- Following769
- Followers17.8K
- Likes105K
Pinned Tweet
I built Universal Prompt Enhancer: a tiny qwen based 0.8B model that turns rough ideas into prompts for image, video, sound & music generators.
Runs locally. Accepts images for edits. Q6_K: ~630 MB.
Weights + 44K examples, released for research.
Try your messiest prompt 👇
I'd give DSH positive points for this rather than penalise. Every other harness leaves so many processes running in bg that I've to ask an agent to clean them up regularly.
DSH is working per expectations here.
You asked for dsh, I added it
DeepSeek Harness is the open-source coding agent by @deepseek_ai. It's now on the @huggingface Harness Leaderboard, which compares ten coding agents on the same open models.
DSH is statistically tied for first in all five evaluations: DeepSeek-V4.1-Flash and GLM-5.3-Flash at low and high effort, and Flint.
It shines on DeepSWE, where agents change code in open-source repositories: 2nd of 10 averaged over the three models, 0.5 points behind Hermes Agent by @NousResearch. On Terminal-Bench's command-line tasks it ranks 6th.
One habit explains much of that gap: when DSH finishes a task, it shuts down every background program it started.
Three Terminal-Bench tasks ask the agent to start a server, for example a model inference API. The grader tests that server after the agent has finished. With DSH, the server is gone by then.
On DeepSeek-V4.1-Flash and GLM-5.3-Flash, DSH passed 2 of 36 attempts on these tasks. @opencode, @pidotdev and Hermes Agent passed all 36.
Those three tasks cost DSH 1.5 to 1.7 points in every evaluation. On DeepSeek-V4.1-Flash at high effort, that is more than its 1.3-point gap to the leader.
Developer preview: 0.2.2-rc.1. Web search is off because it needs a DeepSeek account key.
It costs $0.53 per solved trial, averaged over the three models.
Leaderboard: huggingface.co/spaces/joelni…
1
301
I keep telling this to everyone who'd listen! AI agents are a net positive to defenders.
after running our security agents 24/7 i feel better about where things are going
they find quite things quickly and have access to internal details which help
they should be more effective than an external agent attacker. the asymmetry still exists but not as bad as i thought
15
802
I'm utterly bewildered that I've to debunk these kind of rumours in the age of AI agents being everywhere at beck and call to tell the truth
Incorrect. Telecom / embedded domain engineer for a couple of decades here. And I even specifically worked on antennas on both mobile and network side from 2G onwards.
This claim is utterly false that wired headsets can act as antenna for 4G/5G signals.
In the late 90s/early 2000s, headsets were used as antennas for FM signals for sure. But they are not the right length, not the right impedance, not the right tuning, and not even connected to anything for cellular RF circuitry to be able to act as an antenna for 4G/5G.
4
25
2,057
Looking at the results here, I'm pretty sure the tests aren't setup right. Not to undermine the work done here but evals are hard! Especially when evaluating harnesses, it is extremely difficult to compare what's apples to apples and apples to oranges and everything in between.
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too?
Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model.
V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0.
Results are coming in live. Link below🧵
2
787
The best take on languages in the era of agents!
Rust is a superb language and ecosystem, especially given it was invented pre agentic coding.
With agents, new languages will blossom, even purpose-built ones for specific usecases, and discretionary C or even LLVM IR, assembly…
AI is fluid. No more one-size-fits-all.
17
1,345
Building a sw controller for my vsdinside n4 pro (Deepseek reverse engineering, fable 5.2 designing, opus/sonnet coding).
It needs a lot of physical effort too.
So i pointed my insta360 go3s to it and fable is using it to test results itself.
Need to buy/build a robo arm next!
2
14
789
If you are seeing worse cache hit rates on @DeepSeekHarness compared to @pidotdev I know the reason
While working on my Claude code plugin for DSH, I found a default "tool-result-pruner" plugin they mount
It prunes tool results when context is under pressure, before compaction. This is bad for prefix caching & will increase costs!
No easy way to disable it either with a patch cordis yaml. You'd need to clone a new profile and delete the lines for this plugin from there.
7
28
2,433
My current monthly AI sub stack that is almost endless, with great quality, for most hobby dev work!
- Claude Max ($100)
- ChatGPT/Codex Pro ($20)
- Cursor Pro+ ($60)
- OpenCode Go ($10)
- DGX Spark ($0)
Total: $190 per month.
But I can probably cut out codex/cursor without much loss to get to $110 per month.
Most people into hobby dev can do with just $30 (Claude pro + Opencode Go)
Model Usage:
- Opus 5.5 (Main model for most projects)
- Deepseek 4.1 Flash (Subagent + Reverse - Engineering/Exploration main model)
- Qwen 3.8 Flash Next (Subagent + local sensitive work main model)
- Grok 4.7, GPT 6.1 Sol (subagents, because I have their subs)
- Several image/video/audio models for local usage
If WhatsApp just blocked all the scams (which it can easily do through reports and number checks), we wouldn't have to deal with issues like SIM binding which the govt is now enforcing.
Forgot to mention that not only is this a direct revenue stream bump for OpenAI, but also a big ploy in 2 other ways:
1. Data (More authenticated, authoritative data as properties keep blocking external agents)
2. Ads (Will fuel their ads biz like no other)
This is very interesting!
OpenAI is currently lagging in frontier model race compared to Anthropic, but they have a far better GTM play.
There are a LOT of usecases outside coding and most of these don't need the absolute best frontier model as well.
1
8
1,111
Reads like a prepared PR statement because I don't see much new here as to what WhatsApp/Meta have been trying for years to make it into a "super app" and shoe-horning in AI, but failing.
And they haven't really listened to users whose top ask has been to reduce/remove spam.
I am nearly 4+ Months In WhatsApp
I think of WhatsApp like a shared language, spoken by 3bn+ people. Sometimes we use words, and sometimes it’s a voice note to a parent, a sticker that ends an argument, or a "that was a fun dinner” at the end of the night. Language is at the heart of everything we as people do. We write letters, tell stories, get things done and make plans for the future. A WhatsApp chat is often where we start.
Like any language, the people who use it shape it. They don't wait for permission. They find new things to say, and new ways to say them. Often, our users show us what should come next based on how they’re using the app and communicating in the world.
I’ve spent the last few months listening and understanding what our users want from us and what I heard is simple: help people stay close to those who matter to them, and get more done on WhatsApp. So that's what we'll build.
First, we want to make connections richer.
Sometimes hearing someone's voice, sending a joke, agreeing on a plan or knowing they're safe is the whole point. People send over 10 billion voice messages and over 4 billion stickers on WhatsApp every day. As the way people communicate evolves, we will evolve these experiences with them. We'll make voice notes, groups, stickers and video simpler and more expressive, because how people feel close keeps changing.
Second, we want anything you do on WhatsApp to be as simple as sending a message. Too many things start in a chat and end with many more steps somewhere else: paying a bill, borrowing money, staying connected when you travel. Each handoff to another task costs time. We're working with financial services and telecom partners so that finishing something feels as easy as starting it.
Third, we want to make this all more efficient with AI. It should quietly find what you need, get things done and make sure you never miss a booking or an appointment. The best technology saves you time, so you can get back to the connections that matter.
None of this works without trust. WhatsApp handles over 100 billion messages and 2 billion calls every day. Trust takes years to earn and a moment to lose, and I don't take that lightly. Whatever we build, WhatsApp must remain simple, reliable and private. These principles we’ve had for years aren’t changing. Your personal messages and calls stay end-to-end encrypted. This and our quality bar stays the same.
What happens inside WhatsApp only matters for what it makes possible outside it: more time together, stronger communities, and fewer barriers between people and the lives they want to live.
Onwards.