Gemini 2.5, Deepseek V3, OpenAI image generation "Ghiblifies" the world, Ideogram, Reve and other stories.

Grab your coffee - let's rewind and uncover last week's whirlwind of AI updates and insights.

In this edition

• Google Launches Gemini 2.5 Pro Experimental, its most Intelligent Model Yet.
• OpenAI introduces Native image generation in GPT 4o and Sora.
• DeepSeek releases V3-0324, its latest upgrade to V3.
• Ideograms new model
• Reve emerges from stealth
• ARC-AGI-2 Benchmark
• Other updates
• This weeks AI tools

1. Google launches Gemini 2.5 Pro (Exp) and tops the benchmarks.

Q: Whats the big deal here?

A: Google DeepMind just dropped its most advanced thinking model yet: Gemini 2.5 Pro Experimental. It’s not just smarter- it’s designed to reason, build, analyze, and adapt. It’s topping leaderboards and raising the bar for what AI can actually think through.

Our experiments with this model show we are in wild uncharted territory here. The leap here is quiet significant. As on today, there is no comparable model out there, and Google has convincingly taken the lead here. For how long? That we don’t know.

Gemini 2.5: Our most intelligent AI model

Gemini 2.5 is our most intelligent AI model, now with thinking.

GoogleKoray Kavukcuoglu

Q: 🧠 What makes Gemini 2.5 Pro model a “Thinker”?

Great question! Here's what sets it apart:

• Understands context before responding
• Solves complex, multi-step problems.
• Handles nuance and ambiguity
• Draws logical, informed conclusions
• Uses chain-of-thought to simulate reasoning
• Analyzes information more thoroughly than previous models
• Draws logical conclusions with better context awareness
• Makes more better decisions with improved accuracy

Q: So how good is Gemini 2.5 Pro? 

A: 🏆 Short answer? It’s crushing it.

• #1 on LMArena - tops human preference benchmarks
• Leads in math & science - GPQA, AIME 2025.
• SOTA performance -without expensive tricks like test-time majority voting
• Scored 18.8% on Humanity’s Last Exam - the toughest reasoning benchmark around
• Excellent at coding - dominates SWE-Bench Verified with 63.8% using a custom agent.

Q: 💻 Can it code?

A: In fact, Coding Skills are one of its superpowers.

• Generates full-stack web apps from a one-liner
• Excels in code transformation, editing, and reasoning
• Supports agentic workflows (i.e., can use tools and multi-step plans)
• Great at building games, bots, and apps with minimal prompting

Q: What else makes Gemini 2.5 special?

A: It’s not just smart-it’s also spacious and multimodal.

• 1M token context window (2M coming soon!)
• Multimodal native - handles text, code, images, audio, video
• Massive context comprehension - perfect for documents, data, codebases, and more.

Q: 🚀 Can I try it now?

A: Yup. It’s live and ready for experimentation. Where to access Gemini 2.5 Pro:

• Google AI Studio
• Gemini App (Advanced users only for now, Now available to free users too)
• Coming soon to Vertex AI

Q: So what have people been using it for?

A: Here are some use cases.

• Tell it to create a basic video game. It one shots it by applying its reasoning capability to produce the executable code from a single line prompt.
• Create a simple 3d car simulator with Three.js
• Create a freeform drawing app in SwiftUI
• Use it for creating Complete SaaS applications.
• AI agents that automate your workflow

Q: So what about the pricing of this model?

A: Google will introduce pricing in the coming weeks, which will enable people to use Gemini 2.5 Pro with higher rate limits for scaled production use.

2. OpenAI Introduces Native image generation in GPT 4o and Sora 

Q: What did OpenAI announce ?

A: OpenAI just dropped native image generation directly into GPT‑4o. No more switching to DALL·E-it’s all built into the same chat. And it’s good. Like, finally worth using good.

In a typical OpenAI style, they announced a livestream on X , got together on a table, and showed some incredible generations. Heres the recap: 4o Image Generation in ChatGPT and Sora .

GPT‑4o image generation isn’t just an upgrade. It’s a turning point. Visual content now feels top notch feature in LLM workflows - editable, controllable, readable, and actually useful.

This isn’t just AI that paints-it’s AI that designs with purpose.

Q: And What happened after that?

A: Do you see a torrent of Images in the style of Studio Ghibli right now? Blame the incredible new feature in GPT-4o and Sora. Not many observed this but, They carefully guided the experimentation by users with a demo of a Studio Ghibli style generation during the livestream.

It took X(Twitter) by Storm and everyone was creating Studio Ghibli style images. People loved it so much that OpenAI GPUs were “melting” and they had to put rate limits on image generation.

ChatGPT is now is an environment where we could chat with ChatGPT and create precise visuals as part of the conversation.

Q: 🧠 So what’s actually new here? 

A: GPT‑4o’s image generation is a huge step forward. Instead of duct-taping a text model to a separate image model, this is fully integrated, autoregressive, and context-aware.

• Truly multimodal – Text and image are now co-equal citizens
• Text you can read – Diagrams, infographics, and UI mockups now render cleanly
• Conversational editing – Tweak images using natural language
• Better object handling – Up to 10–20 distinct elements per prompt
• Maintains consistency – Characters, styles, layouts remain stable across edits

Q: How does this compare to DALL·E?

A: DALL·E was a charming try. GPT‑4o is the real deal. Why GPT‑4o is a Big Upgrade:

• No disjointed handoff between models
• Infinitely better text rendering
• Tighter control over details and layout
• Feels like chatting with a designer-not wrangling a paintbrush.

Q: What kinds of images does it handle well? 

A: It’s strongest at useful and informational content - less trippy, more real-world tools. Image Types GPT‑4o Nails:

• Menus, posters, invitations
• Infographics, scientific diagrams
• Product shots, mockups, UI/UX wireframes
• Scene compositions with strong prompt grounding
• Multi-turn refinements (design > tweak > finalize)

Q: Is it photorealistic? 

A: Yes. It can mimic DSLR depth of field, sketch styles, watercolor, risograph, 3D renders, and even chaotic paparazzi aesthetics.

Q: What about limitations? 

A: Yep, there are still a few known issues:

• Sometimes crops long images poorly
• Struggles with non-Latin scripts
• Can hallucinate tiny text or dense info
• Edits aren’t always surgically precise
• Faces in uploads may not remain consistent
• Maxes out around 20 distinct objects

Q: What’s under the hood? 

A: GPT‑4o treats pixels, text, and sound as part of a joint distribution, trained using a single transformer. That means tighter post-training, world knowledge infused visuals, and a shared reasoning stack between modalities.

3. DeepSeek Launches Powerful V3-0324 Model

Q: Wait, what did DeepSeek just launch?

A: DeepSeek dropped a new version of its model-V3-0324-and while it was initially positioned as a minor update, it’s turning heads with real improvements under the hood.🚀

Q: What’s new in V3-0324? Why should I care?

A: Here’s what makes this model worth your attention:

• Huge but manageable: A 641GB model that runs on high-end PCs
• Smarter architecture: Uses Mixture of Experts (MoE), activating only 37B parameters/token-efficient and powerful
• Skills upgrade: Big leap in math, frontend design, and coding
• Open-source: Released under the MIT license
• 'DeepThink' turned off for now-but you can still play around with the rest
• Try it out: Fully testable on DeepSeek’s platforms
• Performance: Scored 55% on aider’s polyglot benchmark.
• Ranking: #2 among non-reasoning models (just behind Claude 3.7 Sonnet)
• Signal from China: A clear sign of China's growing presence in open-source AI
• Dev-friendly: Powerful but lighter compute requirements-ideal for individual developers
• Open vs. closed: A strong open-source alternative to commercial models like GPT-3.5 or Claude

Q: Sounds impressive-but so what?

A: Yes, here’s why this Matters:

• Closes the gap between open-source and proprietary models - especially in complex tasks like coding and math.
• Enables more people to experiment with powerful models without needing cloud-scale infrastructure.
• Signals acceleration in China's open-source AI ecosystem-more competition means more innovation for everyone.
• Sets the stage for upcoming releases (DeepSeek R2, V4), which could further shake up the landscape.

Q: Can I try it right now? 

A: Yup! Head over to OpenRouter or HuggingFace and take it for a spin. Play around with frontend code generation, math problems, or anything else you'd usually throw at a GPT-style model.

Q: Where can I read more?

Hop on to the DeepSeek API docs.

Read some more here: https://simonwillison.net/2025/Mar/24/deepseek/

In Other updates

Ideogram Unveils Groundbreaking Ideogram 3.0

Just when we were absorbing the magnitude of OpenAIs GPT 4O native image creation, in a stunning leap forward, Ideogram launched Ideogram 3.0, a free, state-of-the-art image model with advancements in realism and creative design.

• Unmatched Photorealism and Performance: Ideogram 3.0 blurs the line between generated and real imagery, achieving an Elo ranking of 1132 outperforming competitors like Imagen3, Flux Pro 1.1, Recraft3, and Dalle 3 as evaluated by professional designers.
• Advanced Text Rendering and Language Understanding: The model excels in handling complex text and prompts, enabling precise, stylized designs and intricate compositions with sophisticated lighting, backgrounds, and colours, as seen in surreal underwater and misty forest scenes.
• Style Reference Feature: Creators can upload up to three reference images to guide the style of their generations, offering a faster and more expressive creative workflow for specifying hard-to-describe aesthetics.
• Random Style and Style Code: Users can explore unique styles from a library of 4.3 billion presets using the Random Style feature, with the ability to save and reuse preferred styles via a Style Code for consistency across projects.
• Magic Prompt for Professional Designs: With minimal prompting, Ideogram 3.0 generates professional-quality logos, posters, ads, and landing pages. Examples include four polished logo, ad, and landing page ideas for a fictional coffee shop, “Brewgram,” created in seconds.
• Batch Generation and Scalability: Teams can customize graphics at scale and ideate quickly, reducing traditional design costs and time, making it ideal for marketers and creatives. Ideogram 3 is available for free at ideogram.ai . the API access is coming soon.

Reve emerges from Stealth - tops Image generation benchmarks

What happened?

A: Reve emerged from stealth with Reve Image 1.0 offering impressive prompt adherence and long text rendering-with natural language editing! it was released in the Benchmark Arenas under codename “Halfmoon“.

The model ranks #1 in Image Arena, outperforming Imagen 3, Midjourney v6.1, and Recraft V3

Why it matters:

• Exceptional prompt accuracy, text rendering, and photorealism
• Strong natural language editing, long-text rendering, and intent alignment
• Built by former Adobe & Stability AI leaders with a focus on logic-driven, intent-aware generation
• Try it: Preview is live now - no API yet, but more is on the way.

ARC-AGI benchmark’s new version

ARC Prize Foundation launched ARC-AGI-2 - a set of tasks humans find easy but AI struggles with! The foundation's 2025 competition launches soon with a $700k prize for the first to achieve an 85% score within cost limits. The current best is o3-low's 4%

Figure AI showed off a humanoid robot that walks like a human after hours of training.

OpenAI’s voice assistant got a big upgrade-now more natural, with new personalities.

Perplexity AI added rich media and shopping options to search, making results more actionable.

Anthropic released a "think" tool for Claude. Also, they revealed how its Bot - Claude processes information, explaining internal mechanisms for multilingual reasoning and advanced planning.

This week’s AI Tools

Heygen - heygen.com - AI avatar creation platform for generating custom avatars for digital content and video applications.

NotebookLM - notebooklm.google - Google’s AI tool that transforms notes, papers, and docs into an interactive podcast and knowledge network for better learning.

Jasper - AI Copywriter - Jasper.ai - AI Content Generator for Teams. Jasper is the AI Content Generator that helps you and your team break through creative blocks to create amazing, original content 10X faster

Excuse Generator - Excuses AI - Use AI to generate the perfect professional excuse. 😉 

That’s a wrap for this week in AI - stay curious, build boldly, and we’ll catch you next Week. For more visit our website https://technocient.com