David Fields retweeted
There's a skill for that!
Loop Design Check, an agent-loop skill by affaanmustafa from ECC. It covers goals a machine can check, plan/build/judge loops, and keeping an agent from gaming its own verifier.
Qwen 3.6 27B + skill: 93%
Qwen 3.5 9B + skill: 81%
Claude Opus 5.5, no skill: 76%
Same 8 questions for every model, 2 runs. Full card:
ss33.ai/skills/affaan-m/loop…
David Fields retweeted
Hmm! 'There's a Skill for the that'
"Skills turn open models into frontier intelligence. Find yours at ss33.ai."
ReadyAI subnet 33
powered by Bittensor
Yesterday we launched the store. Today, the proof.
Same job, same brief. An open model with the right skill against Claude Opus 5.5. On 95 skills tested so far, the open model wins.
Skills turn open models into frontier intelligence. Find yours at ss33.ai.
An open model + one skill beats Claude Opus 5.5.
95 skills tested so far beat Opus 5.5 with an open model. 26x cheaper. 6x faster.
Hundreds more improve the open models significantly and we are just getting started.
Skill Store 33 is live. Find the skill for your job and load it into your agent in one line.
ss33.ai
David Fields retweeted
Skill Store 33 is live ss33.ai
Open models can do frontier work. They just need the right skill.
61,000+ skills your agent can search by task. Hundreds already verified to outperform frontier models alone. SN33 keeps the verification pipeline moving.
Why we built it, how we test skills and the roadmap:
PHP2Shell is a critical remote code execution flaw in PHP, and most AI models had never heard of it. They were trained before it existed.
So we tested it. The same models that missed it found it once they had one SN33 security skill.
That's why we're building Skill Store 33 with @ReadyAI_. Your agent is only as current as the skills you give it.
PHP2Shell hit the tech world like a ton of bricks. And everyone had the same question: Is my system vulnerable?
Here's the catch. If you ask your AI agent, the model behind it was probably trained before PHP2Shell existed (we checked!), so it doesn't know what to look for.
The risk sits in the PHP runtime and reaches across your stack: mod_php, PHP-FPM behind Nginx, php-cgi, plus your network and firewall setup. Knowing whether you're exposed means understanding how those pieces fit together, and that's where traditional scans fall short.
An LLM with the right skill can connect the dots. In our tests, models that missed PHP2Shell found it once they had one SN33 security skill. Across 20 audits of PHP apps, that skill cut missed flaws by 41%.
Give it authorized access to your server and it can review the config, check your exposure to CVE-2026-109618 and CVE-2026-109619, and put the real fix first: upgrade PHP and restart every worker.
Care about security? Skill up on SS33, Skill Store 33, coming soon.
I've watched a lot of AI security demos that only work on the app they were built for.
So we tested this skill on an app it never saw. GPT-6 Luna went from 19 of 33 known flaws to 32 of 33.
Skills make cheap models, frontier models...available on the Skills Store later this month 👀
3D scenes used to be frontier-model work.
Here a small model does it with one @ReadyAI_ skill: same brief, same model, and the only change is the skill. Under 2 cents a scene.
Small models with the right skill keep closing the gap. What scene should it build next?
This week I tried to get frontier models to help with basic defensive security work. Blocked at every step.
Even this trailer's hook tripped a safeguard.
So we gave small models one SN33 skill. All 4 found more flaws. DeepSeek: 87% for 2 cents.
What should it audit next for the Skills Store?
David Fields retweeted
Getting frontier models to help with security work is harder than it should be. Even the hook of this trailer tripped a safeguard.
Meanwhile a 2-cent model with one SN33 skill found 87% of the known flaws in our test apps.
5 apps, 4 models. Skill Store 33 is coming...
SN33 skill with Deepseek 4.1 Flash better than SOTA Claude Dashboards for 26X cost and 6X cheaper
The future is open-source + custom skills
Same job as Claude’s new dashboards (with 10M views):
Claude w/ Opus 5.5 vs DeepSeek V4.1 Flash with one skill from ReadyAI SN33.
Watch both and judge the quality yourself.
Flash with SN33 Skill: 26x cheaper. 6x faster.
This is just the beginning of the Skill Store disruption with @ReadyAI_
David Fields retweeted
Subnet 33.
Presentation skillz for agents.
powered by Bittensor
The presentation skills we're building on SN33 @ReadyAI_ keep getting better and better.
Latest: Prezi-style zooming decks. Average score 28 → 98 with the skill vs without, across 60 decks, 3 models.
Games and presentations so far. New domain next week. Store launches later this month.
The biggest problem with skills today is you don't know if they are effective. SN33 skills are all tested and verified.
This slide-deck skill from the subnet took two budget models from an average 24 to 94 across 150 decks. A version frozen before we wrote the test briefs scored the same.
Most skills won't do this. Skill Store 33 shows which ones do. Coming later this month...
David Fields retweeted
SN33 is closing in on six-figure ARR from the jobs API 📈
Last week's pace is now $95k. October is set to be the month we cross it.
September buyback just completed and it was 45% higher than August.
Extrinsic:
taostats.io/extrinsic/922141…
Heads down on Skill Store 33 for @ReadyAI_. The results keep getting better.
Same model. Same prompt. One shot.
Without skills: it plays, but you can't read it and visual quality is poor.
With SN33 skills: readability scores 100% and quality pops off the page!
Last round, a budget model with skills matched the flagship at 1/6 the cost.
Sharing results from our next domain later this week 👀
Grateful for the SEC putting forward clear rules and guidelines that have gone unresolved for too long.
One of the main items we have pushed for is clear guidelines on token buybacks for protocols like bittensor:native and its subnets. We now have them.
Great work @HesterPeirce!
Honored to meet with @HesterPeirce to discuss @ReadyAI_ and broader policy needs for crypto/AI protocols like Bittensor.
Thank you for your continued support of builders and clear, workable rules for the industry to keep it in the USA 🇺🇸
David Fields retweeted
Skill generation through the subnet available now: api.readyai.ai/jobs-api/docs
Skill Store 33 coming in October 🫡
Last week I said 670,000 agent skills have no trust layer. This is what one looks like.
Same model, same prompt. A bot played both builds and ran 14 checks.
No skills: 3/14.
SN33 skills: 14/14.
Skills that just work shipped with proof. Coming to Skill Store 33 👇
From broken to polished: what SN33 skills do for a one-shot game build 🎮
We gave a coding agent a one-page spec for Deregistration Run, a 3D arena where you mine ore against 11 AI rivals. One prompt. No human edits. Then a headless bot played the result and ran 14 gameplay checks: rules, emissions, deregistration, balance, persistence.
Same model (gpt-5.6-terra). Same prompt. Same sandbox. The only difference was whether SN33 skills were installed.
The agent was never told the skills were there. It found them, picked the ones it needed, and read them before writing a line of code.
Without skills: 3 of 14 checks pass. The input API throws, so the test bot can't play it. Half the playfield is lost in the dark, and the miners are placeholders on a flat grid.
With SN33 skills: 14 of 14. It plays through epochs, deregistration and restart, with a lit arena, distinct ore tiers and a dynamic camera.
The cost math 💸
• gpt-6-astra (flagship), no skills: 14/14, $4.12 per game
• gpt-5.6-terra + SN33 skills: 14/14, $0.66 per game
• gpt-5.6-terra, no skills: 3/14, $0.27 per game
The cheaper model matches the flagship's score at about a sixth of the cost. Skills are the difference between a broken build and a game that plays.
