NOVA

AI Video • Tech Tools • Future Media
Exploring how AI is reshaping creativity and digital skills . DM for Collaboration

[NOVA @TechWith_Nova]

Warner Bros and Netflix executives just froze a $100,000,000 historical epic production after seeing this 48-second clip.

A 20-year-old solo creator generated a hyper-realistic scene of a time-traveling travel vlogger interviewing King Leonidas at Thermopylae in 480 BC with zero CGI budget.

Total setup cost: $25.
Render speed: 4 minutes.
Film studio hired: NONE.

142 million views in 3 hours. Half the comments are arguing over the historical accuracy of the 300 Spartans, the other half are demanding a full documentary series.

Here is why your brain falls for this visual magic:

1. Dynamic Lip-Sync & Facial Physics: Perfect voice alignment, realistic sweat textures, and organic micro-expressions on both the presenter and Spartan soldiers mid-march.
2. Multi-Agent Cinematic Consistency: Continuous physics rendering of brass armor reflections, dust particulate dynamics, and synchronized footstep impacts.
3. Spatial Camera Logic: Seamless hand-held vlog aesthetic blended with Hollywood-grade depth of field and realistic outdoor light refraction.

The dynamic next-gen AI stack operating under the hood:

-> Claude 3.7 Sonnet for historical dialogue script, spatial camera vectors, and narrative pacing
-> Veo 3.1 for rendering hyper-realistic cinematic lighting, mountain environments, and volumetric dust
-> Kling 3.0 Omni for generating multi-character marching mechanics, shield physics, and lip-sync alignment
-> Seedance 2.5 for locking facial identity, armor detail consistency, and historical costume accuracy
-> ElevenLabs Spatial for 3D binaural shield clashing, marching footsteps, and acoustic voice reverberation
-> CapCut Pro for adding raw mobile camera noise, dynamic speed ramps, and historical color grading

Hollywood spends hundreds of millions flying 200-person crews to Greece with heavy camera rigs and extra actors.

This creator rendered a full-scale ancient battle documentary before her morning coffee got cold.

[NOVA @TechWith_Nova]

@futurebylea Catching failures early makes all the difference

[NOVA @TechWith_Nova]

@AIwithGhotai A win-win for both existing and new Dreamina users

[NOVA @TechWith_Nova]

@ajaysharma_here This can save creators a lot of time on product ads.

[NOVA @TechWith_Nova]

@ajaysharma_here Claude sparked the conversation, puppyone took it further.

[Dhaval Makwana @heyDhavall]

agents are getting pretty good at research.

the bottleneck is still finding the right context on the person.

Ploid Deep Research gives Claude, Codex, or any agent a much deeper footprint to work with across accounts, media, and platforms that are easy to miss manually.

that could make a lot of research workflows much sharper.

[Hasan Toor @hasantoxr]

If you're building agents on top of PDFs, invoices or scanned documents, pay attention to this.

Nace says Drex beats GPT-6 Astra, Claude Opus 4.7 and every other parser they tested, and gets 38% faster corrections on RealDocBench.

RealDocBench is built from real regulated documents, not clean academic PDFs. That's the benchmark you want to see.

Worth testing on your own documents.

[Kalsoom (ghotai ) @AIwithGhotai]

Already got CapCut Pro? There's an exclusive reward bundle waiting for you from Dreamina and CapCut! 🎁

Try it out right here:

https://t.co/V5FhRDA4G2

New to Dreamina? You can score up to 10 free Seedance 2.0 generations, grab bonus credits, and even get a month of Basic for as little as $0.29. The best part is that every single eligible user walks away a winner.

#DreaminaPartner

[Lea Arden @futurebylea]

One thing I’ve noticed with AI agents is that getting them to work is one challenge, but getting them to work consistently in real-world situations is another.

That’s why I find this approach interesting. Testing different scenarios before users run into problems is a big step towards more reliable AI agents.

[Future Coded @future_coded]

Real learning means the model itself gets better and improved, not just the notes around it.

Continual-ARC from @letsunifyai is a clean and important signal. The agents could store conversations, open skills and keep programs, yet almost nothing transferred. A small 9B model whose weights were actually updated both remembered the recurring puzzles and got meaningfully better on new ones.

This is the kind of result that quietly moves the field forward. True on-the-job learning will likely require the harness to train the model, not just accumulate markdown. Benchmarks like this make that path clearer.

Grateful for the open paper, code and leaderboard.

The next wave of agents will be stronger because of work like this.

[AIMATRIX @AIScout22]

Most "self improving agent" claims rest on one assumption: that a growing pile of notes is learning.

Continual-ARC tests that directly. OpenClaw, Hermes Agent and Prime Agent all solve 93–95% of puzzles, but none get cheaper as puzzles return. Remove their memory and nothing changes.

Training the weights of a small 9B model did what notes couldn't: 57% less spend on puzzles it had met, 29% less on 100 new ones.

For those of us building agents, the harness should be what trains the model, not a folder of markdown files.

[AIMATRIX @AIScout22]

The gap between "works in my demo" and "breaks in production" is where most AI agents live right now.

Plurai Simulation closes it by giving your agent a world to fail in safely: tools, environments, and realistic users all rebuilt from your code, so you catch edge cases before your users do.

Built by people who shipped IntellAgent. Worth a look if you're running agents past the proof-of-concept stage.

Free to start: https://t.co/mLs8Iuwtde

[Kshitij Mishra | AI & Tech @DAIEvolutionHub]

Building an AI agent that works is impressive.

Building one that actually learns from experience and gets better over time is a completely different challenge.

The real question isn't whether your agent works.

It's whether it's improving.

[NOVA @TechWith_Nova]

@AIwithArsalan H3 Image makes creating these kind of technical visuals feel much more accessible.

[NOVA @TechWith_Nova]

@commonman_16 Completely agree that proper attention on AI memory is overdue. The results here make it obvious why.

[NOVA @TechWith_Nova]

@AIwithGhotai The ability to verify fixes is what stands out to me here

[Kalsoom (ghotai ) @AIwithGhotai]

This is the kind of testing AI agents actually need.

A successful demo doesn’t necessarily mean an agent is ready for real-world use. The unexpected inputs, edge cases, and user behaviour are where things get interesting.

Being able to simulate those situations, identify failures, and verify fixes before deployment is a genuinely useful approach.

Love seeing more focus on making AI agents reliable, not just impressive in demos.

[Jaynit Makwana @JaynitMakwana]

crazy.. n8n Labs just open-sourced an AI agent with its own computer and browser

i've always said i'd never give an AI agent my inbox, and Nod is the first one that made me rethink that.

we all have a pile of small jobs that never get done. chase that invoice, cancel the trial before it renews, answer the slack thread from tuesday. chat tools explain how, and you still end up doing them yourself.

Nod has its own computer and browser, so it does the errand itself, in Gmail, Calendar, Slack, Notion and on websites with no integration at all.

the jobs i'd hand it first:

→ prep me before every call
→ draft the follow-up in my voice and wait for my ok
→ cancel the subscriptions i pick and ask for refunds
→ download month-end invoices from portals and file them

the reason i'd trust it with those:

→ only your words can give it permission, so a web page or an email can't
→ it checks with you before it sends, pays, posts or deletes anything you didn't ask for
→ the AI never holds your keys, and every payment with a saved card waits for your yes
→ in their WhisperBench test, none of 216 attacks planted a false memory in Nod

it also remembers months of past chats (93.2% on LongMemEval) and drafts in your voice by checking the emails you've sent.

Apache 2.0, runs on your Mac, Windows PC or own server, and you bring your ChatGPT plan, an API key or a local model. every task shows what it cost, and you can set a monthly limit.

it's a research preview, but if you've been waiting for an agent you can trust with real work, this is the one i'd try first.

https://t.co/QnaLKyCJCP

[Kalsoom (ghotai ) @AIwithGhotai]

A client replied to a review email with six notes. In YouArt Desktop each one lands on the shot it's about and the agent works down the list. I didn't have to copy anything over.

[Hasan Toor @hasantoxr]

Dario Amodei and Demis Hassabis think cancer and Alzheimer's disease will be cured in 10 years.

Founder of NovaAtom, @QuanquanGu predicts it would only take 4 years.

He left ByteDance Seed in June, where he built SeedFold, the model that beat AlphaFold 3 on FoldBench.

Now his 15-person team just launched NovaAtom-Lite.

> 79.07% on FoldBench antibody–antigen, vs. 75.58% for IsoDDE.
> 77.06% on protein–ligand, vs. 75.99% for IsoDDE.
> Live API today.

The next frontier is biology.

[Hasan @Ubermenscchh]

capcut pro users, have you checked out the rewards on dreamina yet? 👀

there’s a 100% win rate event for eligible users.

i got 98% off a basic monthly membership for just $0.29. plus 10 free seedance 2.0 generations with no duration limits and extra bonus credits.

definitely worth a look.

#DreaminaPartner