Sora 2 is here Our latest video generation model is more physically accurate, realistic, and more controllable than prior systems. It also features synchronized dialogue and sound effects. Create with it in the new…

The video gen race is heating up fast.

openai.com/index/sora-2/

👉 Lmk what you think.

(4/4)

OpenAI also paired Sora 2's technical gains with thoughtful design: the Sora app includes enhanced privacy controls, like the ability to delete any video of yourself, while centering the experience around social interaction.

(3/4)

More importantly, the Sora app is powered by OpenAI’s newly released Sora 2 model, featuring two huge leaps in generative video:
- Synced audio → speech and sound that finally match the video
- Sharper, smoother quality → realism that feels production-grade

(2/4)

On the tail of Meta, OpenAI just launched a new social app for generative video. Meet Sora — another TikTok-style AI feed, but one where you can create videos of yourself and your friends.

(1/4)

Introducing Claude Sonnet 4.5 Claude Sonnet 4.5 is the best coding model in the world, strongest model for building complex agents, and best model at using computers.

Check it out here 👇

www.anthropic.com/news/claude-...

(5/5)

Why Sonnet 4.5 matters:
- If your product involves AI agents writing or reasoning about code, Claude just became the SOTA baseline.
- If you’re deploying AI into high-stakes, specialized domains, Sonnet 4.5 is now a serious contender.

(4/5)

Sonnet 4.5 shows ~9 point gains in comprehension + reasoning across finance, law, medicine, and STEM.
- Great news if you’re building in technical verticals.
- This is on a non-standardized benchmark, so external direct comparisons can't be made. Test it and see.

(3/5)

Sonnet 4.5 eclipsed GPT-5 Codex in coding performance:
- 77.2% accuracy on agentic coding tasks (vs. 74.5%)
- 82% accuracy with increased compute
- For Claude Max: this beats Opus 4.1’s coding ability — at 5x lower cost.
- Still 2-3x more expensive vs Gemini 2.5 Pro

(2/5)

Claude Sonnet 4.5 just leapfrogged the field. It’s the leading coding model today — and additionally brings large reasoning improvements in finance, STEM, and law.

(1/5)

If you are one of the people saying "well, I wouldn't scroll AI content for hours", I agree with you. But answer this - what about a generation of children who already use TikTok daily?

Its obvious where this is going. Social media should use algorithms to amplify people, not replace them.

(4/4)

The most prevalent form of social media is 7-second videos served by AI. If companies can generate those directly, they unlock:
1. Hyper-personalized videos to maximize engagement
2. Infinite creator-side scale
3. Perfectly crafted ads

(3/4)

Newsflash - there is a reason that:
- Meta offered $100m signing bonuses, before announcing a partnership with Midjourney in August
- TikTok's owner, Bytedance, keeps releasing SOTA generative video models, with the latest being Seedance 4.0
- Grok app already has a feed of generative posts

(2/4)

I hate the direction social media is going, but its so obvious.

On Friday, Meta released a new platform called Vibes - a TikTok/Reels short form video feed featuring entirely generative content, accompanied by a brand new generative video editor.

(1/4)

Google Co-Founder Sergei Brin said AI models perform better if you threaten them. Not gonna lie, I'm not sure I agree 🤣

Gemini CLI: your open-source AI agent Free and open source, Gemini CLI brings Gemini directly into developers’ terminals — with unmatched access for individuals.

Check it out ⚡
blog.google/technology/d...

💰 Google's BEST AI-coding model is now FREE.

Multi-tasking, SOTA results, zero cost.
Gemini 2.5 Pro is now wrapped in a free and lightweight CLI.
1k requests per day and 60 requests per hour @ 1 million context.

No excuse for devs not to try this.
Link in comments.

#ai #dev

Terminator 3: Rise of the (Coding) Machines The reign of each top AI-coding model is now shorter than a TikTok trend.

(4/4) I wrote about what this feels like from the inside:
in real time, the shifting role of engineers in an AI-driven world

📝 Terminator 3: Rise of the (Coding) Machines
medium.com/@eric.r.eliz...

(3/4) What’s happening in the industry?

Startups are swapping models weekly.
Big companies are still deciding what to use.

🔒 For large companies, committing to a model now feels like marrying a cloud provider. Choosing one and locking to it is hard when models drop weekly.

(2/4) What does rapid AI improvement mean for devs?

Scoping clearly.
Designing architecture.
Debugging ruthlessly.

More and more, you need to think about workflows, product, and the big picture.
Routine can be abstracted.

(1/4) Each new AI coding model now reigns for less time than a TikTok trend.

We used to get one big release a year.
Now? Claude 3.5 → GPT-4.1 → Gemini 2.5 → Claude 4 — all in six months.

📉 Performance is skyrocketing, and better models release within weeks.

Sam Altman Sam Altman

Sam Altman's full post: blog.samaltman.com
Do you agree, or does it concern you?

Startups that master AI will outpace the rest.

Automate what’s predictable. Elevate what’s human. Expel routine to amplify where humans matter most.

Sam Altman just shared his newest vision: check it out in the comments. Do you agree? 👇

4. And Gemini 2.5 Pro is cheaper. Much cheaper.
↳ Gemini Pro 2.5: $1.25 in / $10 out per million tokens
↳ GPT-4o: $10 input / $40 output → 4–10x more expensive
↳ Claude Opus 4: $15 input / $75 output → ~13x more expensive

⚡ Great for startups trying to supercharge developers

3. Gemini Pro 2.5 deeply understands large contexts — and makes sense of them
↳ 90.6 on Fiction.LiveBench @192k tokens (deep comprehension)
  vs Claude Opus 4: 37.5 @120k
↳ 87.8% factual accuracy
↳ 🔎 ideal for scoping product changes, reviewing large PRs

2. Gemini Pro 2.5 is built for developers working with AI
↳ not for teams trying to fully replace them
↳ lower performance in agentic coding without human input
↳ higher performance in collaborative code editing
↳ 🚀 ideal for teams using Cursor, Windsurf, etc.

1. Yesterday, Google released a new version of Gemini Pro 2.5
↳ follows a month of releases from OpenAI, Anthropic, and others
↳ strong in code editing
↳ accurately understands LARGE contexts, like entire codebases
↳ 💰 not just better, but much cheaper

Developing with AI just got 10x more cost-effective.

Google’s Gemini Pro 2.5 is here — and it changes the game for devs building with AI.

Here’s what you need to know 🧵

#thread #ai #google #gemini #swe #coding #startup

(#3/3) Google also released SignGemma: the first open multimodal model for American Sign Language generation and understanding.

Dropped just last week, it's a major step forward for accessibility, gesture modeling, and cross-modal AI.

And yes, it’s open source.

Follow for real more.

(#2/3) Google released Veo 3: an impressive generative video model that outputs 1080p cinematic video with sound.

It handles multi-shot coherence, camera movement, and over 60 seconds of fluid generation.

Some samples? Nearly indistinguishable from real footage.

(#1/3) Google is not behind in AI. You’re just not paying attention.

While everyone’s been watching OpenAI and Anthropic, Google has been shipping some of the most impressive models in the space.

Here are two recent breakthroughs you might’ve missed 👇🧵