DeepSeek quietly drops major update

Read time: under 4 minutes

Welcome back, Superhuman. We’ve come to expect a lot of hype around frontier models these days, with companies teasing us months ahead of their biggest releases. So it’s all the more surprising that DeepSeek went in the complete opposite direction this week, dropping a major update without even a blog post to show for it.

Today’s Insights

• OpenAI’s big shuffle, testing for AGI, and Alibaba’s efficiency breakthrough

• The next slate of frontier models could be coming soon

• Tutorial: How to create perfect lip-syncing videos with AI

• 5 new AI tools to boost your productivity

• News, memes, what’s trending on socials, and more

TODAY IN AI

1. OpenAI shuffles its top leadership: The ChatGPT-maker announced a major executive shakeup on Monday as it tries to steady the boat after last year’s big departures. From now on, we might see a little less of CEO Sam Altman, since he’ll be focusing more on behind-the-scenes technical work and research. COO Brad Lightcap will now be the one landing deals and corporate partnerships, while Mark Chen will take over as Chief Research Officer and Julia Villagra will assume the role of Chief People Officer.

2. Frontier models face their toughest challenge yet: The Arc Prize Foundation just unveiled ARC-AGI-2, a new benchmark that evaluates how well AI can pick up new skills. While humans have scored 100% on the test, top reasoning models can’t even get above 4%. That’s because the test doesn’t evaluate mere memorization; instead, models have to rely on “at least some amount of deliberate thinking” to answer correctly.

3. Chip breakthrough slashes costs by up to 20%: Ant Group, a division of Chinese tech giant Alibaba, came up with a new technique that uses a combination of AI and Chinese-made chips to cut model training costs by one-fifth. The unique architecture performed about as well as Nvidia’s H800, which, while not top-of-the-line, is still powerful enough to be banned from China.

4. New image model comes out of the shadows to top charts: A new image model called Reve Image 1.0 emerged from stealth this week to top the Artificial Analysis Image Arena benchmark. The model seems to do especially well with generating text in images.

PRESENTED BY GUIDDE

Complex ideas? Turn static docs → How-To videos (instantly)

With Guidde, you can turn technical info into captivating How-To videos in less than 5 minutes–no skills needed.

Perfect for HR training, sales pitches, and small business owners:

1. Upload a PDF, slide deck, or screen recording

2. Guidde generates a video + captions and narration

3. Add your logos/branding (if desired)

4. Optimize your content with instant revisions

5. Share your guide in any language

Create your first guide today at no cost.

FROM THE FRONTIER

DeepSeek update kicks off what could be the next season of frontier releases

China’s DeepSeek just shadow-dropped a new version of V3 — the base model that came out before R1 — that weighs in at just 641 GB, meaning it could theoretically run on a high-end Mac.

It seems to be especially good at math and coding, and AI researcher Xeophon even called it the “best non-reasoning model, dethroning [Anthropic’s] Sonnet 3.5.” Plus, it was released with an MIT license, meaning anyone can tweak and adapt it with very few restrictions.

But this might just be the appetizer before the main course: V3 served as the base for R1, the reasoning-focused model that shocked the industry earlier this year. If that trend continues, DeepSeek will likely use this new model to train R2. And we probably don’t have too long to wait, with reports suggesting it could be here before May.

But the US might already have its answer: A mysterious new model called “Nebula” recently jumped to the top of the LMArena leaderboard — and rumor has it that it’s a soon-to-be-released model called Gemini 2.0 Pro Thinking. Google’s own research scientists are winking at the possibility. It looks like we’re headed toward a major showdown between DeepSeek and its well-funded US counterparts.

THE AI ACADEMY

How to create perfect lip-syncing videos with AI

• Go to SkyReels AI and sign up.

• Now select ‘AI Videos’ and go to ‘Lip Sync’

• Upload your source video or image.

• Type in the text that you want the character to say. You can also upload audio.

• Select the desired video speed and click ‘Generate‘.

• You’ll get your perfectly lip-synced video ready within seconds.

• Download and share it with others.

PRESENTED BY CODEIUM

No code? No problem. Now, anyone can code with AI

Windsurf by Codeium is the world’s first AI-powered IDE. It understands your code, predicts your intent, and enhances creativity with powerful AI tools.

Whether you’re a pro developer or a total beginner, Windsurf makes coding effortless + building projects a breeze.

Start coding today at no cost.

AI & TECH NEWS

Everything else you need to know today

📱 Sneak Peek: Some Android users say they’ve gotten early access to a highly-anticipated Gemini feature called Project Astra that can answer questions based on what it sees on your screen or through your camera.

👓 Cyber Specs: Samsung is reportedly working on its own pair of Gemini-powered smart glasses, which are set to come out sometime this year. They’re probably the biggest challenge yet to Meta’s well-received AI Ray-Bans.

📊 Digital Divide: A Wired investigation found that three in four coders say they’ve tried AI, while 17% use it virtually “all the time.” But programmers are almost evenly split on whether they’re optimistic, pessimistic, or agnostic about AI.

🧬 Warp Speed: The UK-based Oxford Drug Discovery Institute used AI-powered graphs to speed up its discovery of new Alzheimer’s drugs by nearly 10 times, helping it study 54 key genes in a matter of days, instead of weeks.

💰 Funding: Seoul-based chip developer FuriosaAI declined Meta’s offer to acquire the startup for $800M, opting instead to go it alone as an independent company. And the UAE vowed to spend $1.4T in AI infrastructure and semiconductors in the US over the next decade.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Substrata*: Analyzes real-time human dynamics to help you sell smarter and close more deals. 

✅ Qodo Gen: Edit, Review, and write quality code that works as intended, with fewer bugs using Agentic AI.

✅ Reworkd: Effortlessly extract web data at scale without writing any code.

✅ Skarbe: Enhance your sales process by capturing interactions, getting follow-up suggestions, and automating conversation tracking.

✅ Eddie AI: Create rough cuts, log A and B roll, and automatically edit multi-cams with AI.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Boost Creative Thinking

Prompt: I want to boost my creative thinking skills through fun and unconventional exercises. Suggest a mix of brain teasers, lateral thinking challenges, and storytelling prompts that push me to think outside the box. Include exercises inspired by famous creative minds like Da Vinci, Einstein, or Picasso. Also, recommend interactive methods such as mind mapping, role-playing, or using random word associations to spark new ideas. Bonus: Give me a unique daily creativity challenge for the next 7 days!

SOCIAL SIGNALS

What’s trending on socials today

👨‍💻 Zero to Hero: OpenAI co-founder Andrej Karpathy built an entire iOS app from scratch in about an hour using ChatGPT and Swift, a programming language he’d never used before. Here’s his ChatGPT conversation log if you want to learn from him.

📖 Logo Lessons: Founder Brett Williams showed how to create graphics in the style of Perplexity’s sleek brand design.

🧮 Physics Fix: Physicist Luis Batalha claims he used xAI’s Grok to “fix” a conversion error on one of Albert Einstein’s equations (which Einstein himself later corrected on his own).

🛋️ DIY Decor: Arcade, which lets users create their own jewelry with AI, is now expanding into home goods. You can even upload a photo of your room and ask the platform to generate a rug with complementary colors.

AI-GENERATED IMAGES

Stone Echoes

Midjourney Prompt: A High-contrast black and white view of the Colosseum in Rome, captured from a slightly tilted wide-angle perspective, emphasizing its grandeur. The sky is soft with a few light clouds. The Colosseum is partially illuminated by sunlight from above, highlighting detailed ancient stone textures with sharp, expressive shadows. The foreground features wet cobblestones reflecting subtle light and adding texture. Deep blacks, bright whites, and detailed mid-tones create a striking monochrome effect --ar 9:16

Acquire new customers and drive revenue by partnering with us

Superhuman is the world’s biggest AI newsletter for businesses and professionals with 1M+ readers and 2M+ followers on socials working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.

🧞 Your wish is my command

Got more feedback or just want to get in touch? Reply to this email and we’ll get back to you.

Thanks for reading.

Until next time!

Zain & the Superhuman AI team