I share insights about AI and automation. Sometimes engineering and technology too. The future belongs to those who build it. DM for collaborations ✉️
Joined March 2012
- Tweets32K
- Following8.9K
- Followers81.7K
- Likes66.4K
Pinned Tweet
A benchmark is only useful when people can understand how the results were produced and verify them independently.
That's why StreamEval is worth paying attention to.
StreamEval introduces a leaderboard for interactive streaming video generation, helping
researchers and developers compare models through a more structured evaluation process.
The benchmark covers 5 tasks and includes 300 scripted test cases. It evaluates models
across task execution, state consistency, visual fidelity, and audiovisual coordination, using 10 task-specific metrics scored on a 0–100 scale.
This is particularly important because strong visual output doesn't always mean a model can reliably follow instructions or maintain consistency during real-time interactions.
Another important aspect is reproducibility. StreamEval is open-source, and its
implementation is available on GitHub.
Researchers can inspect the evaluation methodology and reproduce applicable parts of the workflow using the public code and data.
The project was developed by a research team with affiliations spanning Tsinghua University,
NVIDIA, Carnegie Mellon University, UC Berkeley, Princeton University, the National University of Singapore, and Reactor.
AI companies and model developers can also submit their API endpoints to have their
systems evaluated against the benchmark.
For researchers, developers, and others tracking interactive video AI, transparent benchmarks make it easier to distinguish measurable progress from impressive demonstrations.
Explore the leaderboard: streameval.ai
What do you think is the hardest problem to solve in real-time interactive video AI?
#StreamEval #StreamEvalBenchmark #InteractiveVideo
🤝 Paid partnership
How does Manta handle massive coding tasks?
The team tested it with 7.4 million tokens of real coding history, covering 31,450 messages of experiments, logs, and bug fixes.
A major test of how much context Manta can actually handle.
ETERNAL STUDIO AI retweeted
Many questioned how well Manta performs under actual load.
So we gave it some: a real 7.4-million-token coding session, logged from a current frontier agent. 31,450 messages of training runs, logs and fixes.
We cut it, and let Manta write the next turn. ↓
Meet Manta, a new long-context AI model built to handle massive amounts of information.
It can process up to 100 million tokens in a single prompt while keeping all the context.
Early access is coming soon. Definitely one to watch!
Introducing Manta, our long-context model.
Everything you give it, it keeps. Manta reads a hundred million tokens in one prompt.
Early access opens soon: empero.org/manta
🧵
ETERNAL STUDIO AI retweeted
Introducing Manta, our long-context model.
Everything you give it, it keeps. Manta reads a hundred million tokens in one prompt.
Early access opens soon: empero.org/manta
🧵
More resources to explore StreamEval:
🌐 Leaderboard: streameval.ai
📄 Research Paper: arxiv.org/pdf/2610.28402
💻 GitHub Code: github.com/StreamEval/Stream…
🤗 Hugging Face Dataset: huggingface.co/datasets/Stre…
📊 ModelScope Dataset: modelscope.ai/datasets/Strea…
Email is becoming more than just a place to read messages.
Sol turns your inbox into a workspace, handling the busywork of finding information, connecting details, and preparing tasks so you can focus on decisions instead of searching everywhere.
Less back and forth. More work getting done.
Something worth keeping an eye on.
Waitlist: hellosol.app
Been trying Sol, and it's changed how I think about email.
Email has always been where work arrives, not where it gets done. Someone sends two lines, and the real job starts somewhere else. Find the old thread, find the doc, pull the numbers, put it together, then come back and reply.
Sol does that part. It finds the tasks and commitments in my email, gathers the context, and gets the work ready for me to review. I approve it or change it.
I think this is how email should have worked all along.
Waitlist: hellosol.app
ETERNAL STUDIO AI retweeted
Been trying Sol, and it's changed how I think about email.
Email has always been where work arrives, not where it gets done. Someone sends two lines, and the real job starts somewhere else. Find the old thread, find the doc, pull the numbers, put it together, then come back and reply.
Sol does that part. It finds the tasks and commitments in my email, gathers the context, and gets the work ready for me to review. I approve it or change it.
I think this is how email should have worked all along.
Waitlist: hellosol.app
We shipped a BIG update to Sol (hellosol.app)!
Sol is for people who have consequential work in their inbox.
The most fun part tho - was watching the bets behind this update play out with early users:
- 100+ new emails a day - is the norm for a LOT of people: founders, execs, sales, recruiters, consultants, ops.
- To make things interesting - every email asks for something different: a decision, an approval, a meeting, a proposal.
- The heavy lifting is in: arriving at the decision, doing the review, picking if/when to meet, creating the proposal.
- Drafting the reply - is merely the outcome of that heavy lifting. That alone doesn’t solve the real problem.
That’s where Sol is different. It works on every email as soon as it arrives, does the heavy lifting each one needs, and prepares it for your approval.
You approve or edit. It learns each time and keeps getting better.
When your inbox is filled with highly consequential work, you need an assistant that’s doing the heavy lifting, continuously.
That’s what we’re building Sol to be - proactive about what you need, 100 times a day, every single day.
Extending early access to this big update ✌️
Gemini understood the assignment. 💀
Claude played it safe. GPT-6.1 Terra couldn’t even fit the whole phone in the frame.
Gemini takes this round. No debate.
Which one are you buying?
Gave Claude Opus 5.5, Gemini 4 Argon and GPT-6.1 Terra one prompt: recreate the legendary Nokia 3310 as if it was made in 2026.
One of them refused to copy it and built its own phone. Gemini went straight for the real thing. GPT-6.1 Terra got a second try and still cut off half the phone. What the fuck, GPT?
Which one would you buy?
ETERNAL STUDIO AI retweeted
finally. claude code can just do outbound natively now
point it at your ICP and it finds the companies, the right person, and a verified work email. you only pay when it actually finds one
27 people, 20 verified emails, $0.58 total.
It's cheap and fast.
we’ve been paying way too much for lead gen at simi lol. congrats @treg_ai
And the best thing is it's Opensource.
🤝 Paid partnership
Introducing OpenSource Clay killer
- 85% cheaper than clay
- Most accurate on people-search bench
- 25x faster
- Run GTM directly in claude code or codex
Try at treg.to/people-search
Repo link below 👇
Claude Code just made outbound easier. Find leads, get verified emails, and save money. All open source. 🚀
finally. claude code can just do outbound natively now
point it at your ICP and it finds the companies, the right person, and a verified work email. you only pay when it actually finds one
27 people, 20 verified emails, $0.58 total.
It's cheap and fast.
we’ve been paying way too much for lead gen at simi lol. congrats @treg_ai
And the best thing is it's Opensource.
If you're planning Halloween product content, CapCut PC is worth exploring. CapCut Desktop Pro is 60% off and Ultra is 40% off, and October 30 is the last day of the Black Friday deal. I tested AI Video with just one product image and generated multiple Halloween-inspired scenes around it. Different settings, different moods, same product. Then I used AI Edit to tweak individual details without starting the whole video over. One image, multiple creative directions. Pretty useful for small business content. tinyurl.com/2v23rn22 #capcutpc #capcutmarketing #capcutblackfriday #capcutsale
ETERNAL STUDIO AI retweeted
Meet Luka Sorrento. He doesn't exist.
At Flayr Labs we've been doing R&D on AI music and video generation, and we wanted to answer one question: can AI make a music video that actually makes you feel something?
So we ran an experiment. Luka's song, "Don't Call Me When the Quiet Gets Too Loud," is fully AI-generated, and so is his music video, start to finish. It puts him in recreations of 25 iconic music videos, from Thriller to HUMBLE. to Otis, as a tribute to the videos that shaped all of us.
What it took:
• Music: Suno ($8/mo plan)
• Stills: Nano Banana, keeping Luka's face consistent across every scene
• Motion: Kling and Veo 3.1
• Lip-sync: sync-3 and Hedra, fed an isolated vocal track
• Generation spend: about $160, plus a share of a $100/mo Hedra plan we use across other projects
• Time: 2 days
By the numbers:
• 34 original music videos studied frame by frame
• 120+ stills generated
• ~175 AI video and lip-sync renders
• 100+ takes thrown away
• ~15 scenes swapped out entirely
• 45 shots in the final cut
What worked:
• Medium shots with clean light and small movements look close to real
• Studying the real footage before generating anything. Scenes built from memory looked nothing like the originals
• Isolating the vocal from the mix made lip-sync dramatically better
What didn't:
• Lip-sync falls apart when the head turns or moves fast
• Editing real frames sometimes kept the original artist's face, so we rebuilt those from scratch
• Logos we removed would grow back mid-render and had to be tracked and painted out frame by frame
• More than half of what we generated never made the cut
Our take: under $300 all in and two days for a fully AI artist, song and music video. It's already watchable, and the quality gap with traditional production is closing fast. But taste, direction and judgment still do the heavy lifting.
The tools generate. A person still has to decide what's good.
Is AI music viable? Did Luka make you feel something?
Watch and tell us.
right sizing but for LLMs. why did this take so long. RouterPlus
AI teams overpay for inference twice: once for models bigger than they need, and again for idle capacity they don’t use. @useRouterPlus is the extended inference team for fast-growing AI companies. We test models and providers on your traffic and get providers to compete for your work.
Make your inference better. Go to routerplus.com
ETERNAL STUDIO AI retweeted
AI teams overpay for inference twice: once for models bigger than they need, and again for idle capacity they don’t use. @useRouterPlus is the extended inference team for fast-growing AI companies. We test models and providers on your traffic and get providers to compete for your work.
Make your inference better. Go to routerplus.com
ETERNAL STUDIO AI retweeted
Introducing Bivia.
A multimodal AI decision model built to complete entire workflows, not just answer individual questions.
We've spent 3+ years building and deploying AI in production across semiconductors, energy, manufacturing, and other complex industries.
Today, we're bringing that experience to everyone.
"⏰ One more thing before you scroll! CapCut PC’s Black Friday deal ends Oct 30, and you can save 60% on Pro or 40% on Ultra.
Here’s a little editing trick I’d use for this floating product effect. Start with the product sitting still, bring in a sound effect as it lifts off, and time the promotional text to appear right after the reveal. 🔊
With Multi-Track Editing in CapCut PC, I can keep the footage, audio, and text on separate tracks, then adjust each one independently until everything lines up. It makes it much easier to control the rhythm of the ad.
That tiny bit of timing can make the whole edit feel more polished. 🎬
tinyurl.com/2v23rn22
#capcutpc #capcutmarketing #capcutblackfriday #capcutsale"
"A practical reminder for anyone planning their next content sprint: there are deals on both creation and editing tools right now.
Dreamina’s first month is $0.29. CapCut PC’s separate Black Friday sale offers 60% off Pro or 40% off Ultra through Oct 30, desktop only.
The pairing is simple: make the visuals with Dreamina, then use CapCut PC to shape them into a cohesive video. With multi-track editing, you can layer footage, product images, voiceovers, captions, and music, then fine-tune the timing and transitions to fit your campaign. Two separate subscriptions for two different parts of the process.
Dreamina: tinyurl.com/5xbtv4x6
CapCut PC: tinyurl.com/2v23rn22
#CapCutDidThat #Dreamina #Smallbusinessmarketing #CapCutSale"