Jared Palmer retweeted
Devin Fusion has claimed a top score of 68 on the @ArtificialAnlys Coding Agent Index with an Opus 5.5 lead and GPT-6 Luna sidekick.
Devin Fusion outperforms Claude Code with Opus 5.5 on both performance and price – earning a score of 68 at a 31% lower cost per task.
Devin Fusion outperforms Claude Code with Opus 5.5 on both performance and price – earning a score of 68 at a 31% lower cost per task.
105
46
412
59,693
One-size-fits-all blanket usage caps on AI spend don’t make sense. On the other hand, yolo tokenmaxxing doesn’t scale either. We’ve come up with a better approach to cost control centered around efficiency. More efficient employees can have limits raised automatically (with a ), while other spend requests go thru an admin review process.
What if your AI budget went to the engineers who use it best?
Blanket usage caps are a hiring freeze in disguise.
Introducing efficiency-based auto-approvals for your most productive engineers, plus Devin Coach to help everyone get there.
devin.ai/blog/efficienc…
22
4
85
8,985
I could be wrong, but from my understanding, without months of weight training, normies will not have the neck strength to tolerate the g force at this speed. You’d would only last a few laps at best before your neck gives out
Tesla Roadster could be 5 s faster than F1 at COTA.
I simulated a lap with the revealed specs: 1:27.9 vs Verstappen's 2025 pole at 1:32.5.
Where it wins: acceleration. All-wheel drive, 2 g of fan suction keeping the tyres planted on exits, and cold-gas thrusters. Top speed on the vs 324. At the end of the lap the temperature in the tank is -150 Deg C.
Where F1 wins: braking and the fast corners. An 800 kg car with downforce that grows with speed beats a 2-tonne one.
As @elonmusk said the tires are the biggest limiting factor, I assumed quite conservative values taking into account the massive vertical forces that they would have to endure.
23
2
90
17,336
We productized how we handle incidents, bug reports, and alerts at @cognition with Devin Oncall. It’s completely changed how we work and also how we think about incident management and SRE. From investigations, to reproductions, to triage, rollout, root cause analysis and even postmortems, we do it all with Devin. Excited to share this with the world.
Introducing Devin Oncall: Devin automatically investigates alerts and bugs, helping your team diagnose and resolve production incidents.
Devin Oncall already handles thousands of alerts every day for some of the world’s largest companies. Let it become your SRE today.
13
4
101
8,838
Visual Basic 6 running in the browser.
Built on top of an OpenAI-like Intelligent/GenUI React DSL + WebGPU.
(h/t @wieslawsoltes)
wieslawsoltes.github.io/VB6/
Built on top of an OpenAI-like Intelligent/GenUI React DSL + WebGPU.
(h/t @wieslawsoltes)
wieslawsoltes.github.io/VB6/
13
9
74
6,983
mise is really good. i switched from homebrew on my work machine and have had a wonderful experience so far
new in mise: tools that keep themselves up to date.
set auto_update on a global tool and mise upgrades it right before it runs. optionally run a daemon to do upgrades in the background.
mise.jdx.dev/dev-tools/#aut…
7
12
298
19,960
Cool write up of how a $200 Kev-4B fine-tune can mog every other decision model.
If you want to do this yourself, there is an official skill:
npx skills add jaredpalmer/kev@kev-finetune
It will discover Jev/kev/decision model callsites in your app and/or interview you about your use case. It then will help you generate synthetic data and walk you through fine tuning on @modal. You can then use the sister kev-deploy skill to deploy your model to @modal endpoints.
github.com/jaredpalmer/ke…
If you want to do this yourself, there is an official skill:
npx skills add jaredpalmer/kev@kev-finetune
It will discover Jev/kev/decision model callsites in your app and/or interview you about your use case. It then will help you generate synthetic data and walk you through fine tuning on @modal. You can then use the sister kev-deploy skill to deploy your model to @modal endpoints.
github.com/jaredpalmer/ke…
finetuned @jaredpalmer 's kev-4b on a mix of internal cyber classifications tasks
pretty non-trivial performance boost compared to jev and luna
x.com/ShrivuShankar/…
24
12
309
20,493
I’ve been experimenting with Devin and autoresearch every night for the past week or so. The prompt I’ve been using for is pretty simple:
`I want you to do autoresearch to improve {verifiable metric(s)}. Hill climb. You have a budget of ${budget} and {time limit}. Use sub-devins to parallelize work. Track your plans and progress as you go in the filesystem. Work methodically. Create abstractions if you need to to help you verify faster. Manage your own PRs/babysit code review and merge as you see fit until you get a candidate that passes gates you set. Boil the ocean. Believe in yourself. You got this. {+more confidence/hype statements}`
After you have the setup you like, just make this into a skill and turn on automatic code review. With this prompt, Devin can run for hours autonomously.
Funny enough, the “boil the ocean” and hype statements seem to yield better exploration behavior and make most models less likely to give up early or reward hack changes/ship incremental fallback stuff. Inspiration for this was @jarredsumner’s prompts during his Riemann work.
`I want you to do autoresearch to improve {verifiable metric(s)}. Hill climb. You have a budget of ${budget} and {time limit}. Use sub-devins to parallelize work. Track your plans and progress as you go in the filesystem. Work methodically. Create abstractions if you need to to help you verify faster. Manage your own PRs/babysit code review and merge as you see fit until you get a candidate that passes gates you set. Boil the ocean. Believe in yourself. You got this. {+more confidence/hype statements}`
After you have the setup you like, just make this into a skill and turn on automatic code review. With this prompt, Devin can run for hours autonomously.
Funny enough, the “boil the ocean” and hype statements seem to yield better exploration behavior and make most models less likely to give up early or reward hack changes/ship incremental fallback stuff. Inspiration for this was @jarredsumner’s prompts during his Riemann work.
24
8
307
21,200
What you're seeing in gaming decompilation, stitching, reverse engineering and then ultimately web porting (via WASM and WebGPU) is a testament to 2 things.
1. Despite all of its limitations, the web is still the best memetic distribution system ever created
2. The Uniform Resource Locator (URL) is one of the greatest inventions of all time. Up there with the wheel, written language, agriculture, printing press, steam engines, transistor, etc.
1. Despite all of its limitations, the web is still the best memetic distribution system ever created
2. The Uniform Resource Locator (URL) is one of the greatest inventions of all time. Up there with the wheel, written language, agriculture, printing press, steam engines, transistor, etc.
There's a reason I've been porting so many classic games I liked to web with WASM and why this is taking off now I think
Because it's way easier to just open a new tab, type the URL and start playing than...
Switching on a console or gaming PC, waiting for it to boot, installing the Xbox/PS/Windows/Steam updates, rebooting, finding the game in your library, then a game popup appears (new DLC!!!), click away popup, click [ PLAY ], game has to download the new 50GB updates for 30 minutes, wait, load game, takes ages, tap [ Online ], try find our friends in some convoluted way, [ Play ]
I think game and console developers do not realize people don't have time to do all this
Many of us are doing other stuff, then have a break, just want to jump into a game, play a bit with friends, and get out again! And decompiling games and porting them to WASM does exactly that!
13
21
504
26,967