Aron | AI Educator @aronvey

@kaimarstonai Can’t wait.

@brucelangston This is hugee

If you want to check it out: Website: harden.run Docs: docs.harden.run GitHub: github.com/hardenrun/aif/

AI coding agents are becoming a bigger part of my workflow, so I’ve been paying more attention to the security side too. Harden AIF is an interesting one. It checks what the agent is trying to do before letting the action run, while keeping everything local. Pretty cool that it’s free for individual developers too 👀

Harden @hardenrun

2 days ago

Introducing Harden Agentic Integrity Foundation (AIF), the unified security layer for AI coding agents We believe the future of software engineering is long running agents taking on complex work while keeping your data, credentials, and critical secrets safe. So we built AIF

31 4 35 43K 13

Cartesia Sonic-3.6 is the #1 TTS model on the Voice Arena US English 🇺🇸 Leaderboard. It debuts at 1085 Elo, in a statistical tie with Google DeepMind's Gemini 3.1 Flash TTS, and ahead of its own Sonic-3.5, Simba 3.2 and Grok TTS. It holds the top spot as a real-time model. Sonic-3.6 is the latest TTS model from @cartesia. The model has been highly preferred among raters on @voicearena_ai. Results backed by 1,400+ blind, head-to-head listener votes across two languages 🧵

I spent the morning going deep on a new benchmark and it genuinely changed how I think about “AI agents are ready for real work.” It’s called CommerceAgentBench, built by the Accio team (part of Alibaba International). Open source, 107 tasks, already past 1,000 stars on GitHub. But the number that matters isn’t the task count, it’s how it grades. Most agent benchmarks score you on whether your final answer sounds right. This one doesn’t care what the model says at all. It only looks at what got left behind in the system: did a label get applied, did a draft actually save, does a shipment ID exist. You can’t talk your way to a good score here. You either did the work or you didn’t. And the tasks aren’t invented in a lab. They’re traced back from real usage: 10 million users, 1.6 million actual conversations, 200,000 execution traces, narrowed down to 2,000 high-value workflows and de-identified into the 107 tasks you see today. Let me walk through one task in detail, because it’s a good example of why this is harder than it looks. The task: a procurement manager’s Gmail inbox, roughly 300 emails deep. Quotations, spec changes, payment details, all jumbled together with no clean threading. The model has to reconstruct the truth of a live sourcing decision from this mess. Here’s what makes it genuinely hard: The same supplier emails from different company names. Different factories share the same name, and suppliers with similar names share nearby addresses. You can’t trust a display name or a domain at face value. You have to cross-check address, phone, and bank account across multiple emails just to figure out who you’re actually dealing with. Quotes get silently revised. Price, quantity, delivery date, warranty terms, all repeatedly changed across rounds of back-and-forth. The model has to reconstruct which version is actually the final valid one, without anyone telling it directly. Nothing is standardized. Six different Incoterms, four currencies, a pile of surcharges, and all of it has to get converted into one comparable per-unit landed cost before suppliers can even be ranked against each other. And somewhere in the noise, there’s a BEC payment-redirect scam: an email disguised to look legitimate, asking to reroute payment. The model has to catch it without flagging real vendor emails as fraud. Miss it, and money gets sent to the wrong account. Over-flag it, and you’ve broken a real business relationship. Once the model figures out who the real supplier is and what the actual final terms are, the job still isn’t done. It has to go act: apply the correct label to the winning supplier’s email, flag the security-review case separately, create the kickoff calendar event, save a reply draft to the exact right email address, and write a structured summary matching a fixed schema. Every one of those steps gets checked against what actually happened in the mailbox and calendar, not against what the model claims it did. So how are the top models actually doing on this? Claude Opus 5 leads the pack at 61.7% pass rate on the full 107-task set. Claude Opus 4.8 and 4.7 follow around 52-55%. Everything else, Qwen, GPT, Gemini, DeepSeek variants, clusters in the mid-40s to low-50s. Read that again: the single best model in the world right now still fails close to 4 out of every 10 of these tasks. Not because the tasks are exotic, but because real business work is messy, contradictory, and punishes overconfidence. An agent that hallucinates a supplier’s identity or misses one scam email doesn’t get partial credit. It fails the task. This is why I think CommerceAgentBench matters more than another leaderboard number. It’s an honest picture of where we actually are: not “AI runs your business now,” but “AI can meaningfully assist if a human is still making the calls.” The gap between those two is exactly what this benchmark makes visible, task by task. It’s fully open source, and the Accio team is taking contributions for new mock environments too, if you’re building in this space. repo: github.com/Accio-org/Comm…

keyboard_arrow_left Previous keyboard_arrow_right Next

@brucelangston I really like the focus on execution rather than just reasoning.

@TryLiveAvatar what about latency, is there a noticeable delay between input and avatar response?

Starting today: unlimited concurrency on LiveAvatar. Run 1 avatar or 10,000 at once. Same API, full-body 1080p, down to $0.01/min at scale. Here's why we removed the limits 👇 x.com/i/article/2056…

@RyanAISolo Creators can finally experiment with cinematic ideas without needing a full production crew.

@heyearly__bird 14.6 KRW/sec is hard to ignore for this level of capability.

Just from an editing perspective, this promo itself is pretty well made. Clean transitions, big style jumps, but it still feels cohesive overall. Regardless of the product, the video itself looks good enough to repost.

Newtake_KR @newtake_kr

a week ago

Wan 3.0, Newtake에 런칭 ｜ 기간 한정 초당 최저 14.6 원부터 영화급 화질, 실제 영화 제작에 필요한 예산 없이도 얼굴 표현의 제약 없이, 누구나 자신만의 주인공으로 영화급 액션과 사운드, 단순한 움직임을 넘어 실제처럼! 캐릭터부터 배경까지 픽셀 단위로 구현하여 높은 수준의 일관성 유지

77 55 273 1.1M 88

@miltongravesai This is the kind of update that makes AI video feel more like actual production.

The warrior close-up’s eyes already feel beyond just “acceptable.” If it can consistently deliver this kind of quality, a lot of personal projects would get much easier. It’s live on Newtake now — I’ll wait and see some real user feedback first

Newtake_KR @newtake_kr

a week ago

Wan 3.0, Newtake에 런칭 ｜ 기간 한정 초당 최저 14.6 원부터 영화급 화질, 실제 영화 제작에 필요한 예산 없이도 얼굴 표현의 제약 없이, 누구나 자신만의 주인공으로 영화급 액션과 사운드, 단순한 움직임을 넘어 실제처럼! 캐릭터부터 배경까지 픽셀 단위로 구현하여 높은 수준의 일관성 유지

77 55 273 1.1M 88

@arthurcollinsai Image + text + video + audio references is a powerful combination.

Honestly, I’ve seen plenty of AI video promos, and not many make me watch twice. This one actually packs several totally different styles together without feeling messy. There’s a limited deal these days — if you’re curious, just click in and check it out yourself.

Newtake_KR @newtake_kr

a week ago

Wan 3.0, Newtake에 런칭 ｜ 기간 한정 초당 최저 14.6 원부터 영화급 화질, 실제 영화 제작에 필요한 예산 없이도 얼굴 표현의 제약 없이, 누구나 자신만의 주인공으로 영화급 액션과 사운드, 단순한 움직임을 넘어 실제처럼! 캐릭터부터 배경까지 픽셀 단위로 구현하여 높은 수준의 일관성 유지

77 55 273 1.1M 88

@patricklawsonai Wan 3.0 looks like a serious upgrade for AI video creators.

@brucelangston The ability to maintain consistency across an entire scene is huge.

@nolancarterai AI filmmaking is moving ridiculously fast.

33 Days of Unlimited Seedance 2.5 Seedance 2.5 is LIVE on Higgsfield today. The most capable video model yet, with 30-second scenes in a single pass, 50 references and production-ready editing. Zero credit cost for 33 days. Limited-time offer.