🍭 DeepSeek Digs Deep 🤖

Good morning. NYT is a sellout, DeepSeek is going deep, and businesses are mostly still scared of AI.

Let’s dig in.

What’s Cookin’:

• More DeepSeek R1 upgrades

• New York Times <> Amazon

• Steal This Prompt: Go Full Warhol

• 5 AI Tools

• And Everything Else you Need to Know

Deepseek
🤖 DeepSeek’s Not-So-Silent Flex: R1-0528 Hits Deep

The Bite: DeepSeek just dropped a spicy upgrade to its R1 reasoning model R1-0528.

And it’s giving Gemini, GPT-4o, and your sweaty MacBook fan some serious anxiety.

This model:

🍭 Hits 87.5% on the AIME reasoning test (from 70%)
🍭 Jumps 10%+ on LiveCodeBench, now nipping at OpenAI’s o3
🍭 Doubles performance on “Humanity’s Last Exam”.
(Which, yes, is a real benchmark, not a Black Mirror episode)

It also introduced a distilled version (Qwen3 8B) that outperforms Gemini-Flash and o3-mini on a single GPU.
That’s like running “Dune: Part II” on a flip phone.

And while OpenAI and Google play API paywall tag, DeepSeek is just out here dropping models on Hugging Face like mixtapes.

Open weights, MIT license, minimal GPU usage.
VCs are sweating.

Snacks: 

• 685B parameters in the new R1 — up from 671B in January

• JSON + function calling support baked in

• Trained in 55 days for $5.58M on 2,000 GPUs

• Can reason, code, and math better than your college roommate

• Comes with a heaping side of censorship — refuses to discuss China’s human rights abuses directly, despite acknowledging them obliquely

• Still more “open” than OpenAI, but less “open” than, say, democracy

Why it bites: This update is a Rorschach test for the AI industry.

→ On one hand: DeepSeek proves that lean, fast, open-source development can hang with (and sometimes dunk on) trillion-dollar labs.

→ On the other: it’s a reminder that “open” doesn’t always mean “free.” When models dodge political critique but still ace logic puzzles, you’re left with a bot that can solve quantum physics — but won’t talk about Xinjiang.

It’s not just a model war. It’s a battle over control, cost, and conscience.

And DeepSeek?
It’s coming for all three.

New York Times x Amazon
📝 NYT Sues, Settles, Sells Out? Amazon’s AI Deal Is Here

The Bite: One year after suing OpenAI for scraping its content, The New York Times just inked a licensing deal with… Amazon.

Yes, that Amazon — the one still trying to catch up in AI like it just logged in.

Now your Alexa might whisper Times articles between weather updates.

Snacks: 

• Multiyear deal includes NYT Cooking, The Athletic, and core news content

• Amazon can train its AI models on Times content (not just use it)

• Terms undisclosed, but comes right after NYT sued OpenAI + Microsoft

• Jeff Bezos owns WaPo, but Amazon still wanted NYT’s editorial clout

• NYT CEO: “High-quality journalism is worth paying for” (finally)

Why it bites: Remember when NYT was waving the copyright flag like a legal Zorro?

Now they’re packaging their archives like a Netflix content bundle — licensing journalism to the same category of AI platforms that turned newsrooms into training data overnight.

It’s a cautionary tale in reverse: sue first, monetize later.

Meanwhile, Amazon has been quietly stacking deals:

Adept, Covariant, Anthropic, and now Amazon.
To build its own AGI Voltron. NYT data might just be the narrative fuel.

And if this works?
Every media company not on a licensing deal better buckle up… or start training their own chatbot named Pulitzer.

🍭 Steal this Prompt
🍌 Pop-artify yourself like a Warhol print

Turn yourself into a Warhol print.

This one’s loud, blocky, and iconic.

Drop in a photo and get back a retro pop-art portrait in bold colors and gallery-grade contrast.

🖼️ Output: Vintage Warhol-style print with punchy color blocks
🧠 Acts like: 60s silkscreen artist + brand stylist
📸 Use it for: PFPs, posters, merch mockups, digital collectibles

1. Hit this link (prompt + references)

2. Upload the reference image first, then your photo second

3. Paste the prompt into GPT-4o and let it cook

🍭 Try the prompt on Snack Prompt 🍭

Together with Recraft
🖼️ Consistent Visual Styles with Recraft

For anyone tired of endlessly scrolling through stock photos and editing them to fit their brand — meet Recraft. It’s an AI-powered image generation and editing tool. The recent update, Advanced Style Creation and Control, makes it much easier to build and maintain consistent visual styles across your work.

Why Recraft?

Trusted by 4 M+ users and teams at Netflix, Ogilvy, and Hubspot, Recraft empowers creators to:

🎨 Explore endlessly with the Infinite Style Library

⚡ Work faster with AI-native tools

🧩 Stay on-brand by saving and sharing styles across teams

From a single prompt and style reference, generate logos, illustrations, icons, mockups, and more, fully aligned with your brand.

Special offer for SnackPrompt’s subscribers: use promo code SNACK11 for a $11 discount on any paid plan!

Big Biz
🚧 Why Businesses Still Hesitate on AI

The Bite: If you’re wondering why your shiny AI pitch keeps getting sidelined, here’s the real reason businesses are hesitating.

Snacks: 3 Invisible Fears

🍭 1. Losing Control
→ They’re terrified AI will go rogue - confidently stating wrong facts, policies, or prices.

🍭 2. No Governance
→ Without clear rules, companies fear chaos—so they freeze.

🍭 3. Lack of Visibility
→ Businesses need oversight. If they can’t see what’s happening clearly, they won’t adopt it.

Why it Bites: Knowing these fears upfront means you can address them proactively before the client even asks.

You can build alone.

But…

You’ll pitch stronger when you’re plugged into a community who knows the objections before they happen.

Small structures.
Clear frameworks.
Big confidence.

That’s why great builders never build alone.

Tool Box 
🧰 Trending AI Tools

🏠 Zown - AI-powered homebuying that pays you back
Smart listings, auto-approvals, offer timing… and up to $25K redirected toward your down payment.

🧠 Circuit Tracer – Anthropic’s AI mind reader
Visualize how LLMs “think” using attribution graphs. It’s like an fMRI scan for neural networks — open source and weirdly beautiful.

💼 Taplio – LinkedIn growth on cruise control
Write posts, engage leads, and track progress — all in one tool that turns vibes into conversions.

🎙️ ElevenLabs Conversational AI 2.0 – Text-to-speech, but make it cinema
Clone voices, narrate stories, or dub your next viral short. So realistic it might ask for royalties.

🎧 Audino AI – Auto soundtrack your content
Drop a video, get dynamic background music and SFX that adapt scene-by-scene. It’s your film school friend… but faster.

Everything Else
🧠 You Need to Know

Google: AI models now run on your phone, offline

→ The new AI Edge Gallery app lets users download Hugging Face models straight to Android — no cloud, no connection required.

VC: “AI’s rise is breaking every tech speed record”

→ Her 340-slide comeback report says AI’s adoption, cost drops, and energy gains are “unprecedented.” 51 slides literally use that word.

Gmail: Gemini now auto-summarizes your long emails

→ AI summary cards will show up whether you want them or not. You can opt out… eventually.

Google: AI Overviews thought it was still 2024

→ A bug made the AI forget what year it was. It’s now fixed, but not before offering some deeply unhelpful timelines.

Meta: 90% of privacy risk checks now handled by AI

→ Human reviewers are being phased out in favor of auto-approvals, even for youth safety and misinformation. Devs cheer. Critics brace.

Referral Program
Invite Friends & Earn Some Freebies 🔍