AI News: OpenAI Dropping Tomorrow! Open Source o3 Level Model, Midjourney V7, and More!

This update dives into the latest buzz in the AI world, heavily featuring OpenAI's plans and recent moves. We get hints about upcoming models like O3, O4 Mini, and potentially GPT-5, alongside the rollout of a significantly enhanced Memory feature for ChatGPT, signaling a strategic shift towards hyper-personalization.

Here's a breakdown of the key topics discussed:

OpenAI Updates & Strategy

• There's fresh information about OpenAI's upcoming models, specifically O3 and O4 Mini, with whispers about GPT-5. Apparently, building GPT-4.5 took two years, so expectations are high for future releases.
• Sam Altman mentioned that GPT-5 is shaping up to be much better than they initially thought, but integrating everything smoothly is proving challenging.
• Demand for OpenAI models, especially image generation, has been off the charts, leading to actual GPU shortages.
• Altman feels the improvements to O3 will be well-received. He can be a bit hypey, but the presenter is personally excited about individual model releases rather than waiting for a combined one, likely pushed by competition.
• Breaking News: Altman teased significant releases happening very soon. Leaked images suggest models like O4 Mini, GPT 4.1, and its variants (Nano, Mini) might be announced shortly.
• ChatGPT Memory Feature: OpenAI launched a major update to ChatGPT's memory.
  • It can now reference all your past conversations (not just explicitly saved memories) to give much more personalized and relevant answers.
  • The presenter agrees with the idea that OpenAI's competitive edge is shifting from pure model intelligence (which is becoming more common) towards deep personalization.
  • This personalization strategy makes sense, especially since recent research (cited from Harvard Business Review) shows AI's top use case is now therapy and companionship.
  • This memory feature uses your past data, creating a unique advantage ('moat') over open-source models that don't have this history.
  • This raises the question of needing an open-source standard for AI memory.
  • You still have control – you can opt-out or ask ChatGPT what it remembers about you.
  • It's rolling out to paid users first, likely followed by free registered users.

Open Source AI Models

• DeepCoder: A new, efficient coding model (14 billion parameters) called DeepCoder was released by Together AI and Agentica.
  • It's considered to be at the level of OpenAI's O3 Mini for coding and reasoning.
  • Crucially, it's fully open source – they shared the dataset, code, and training methods. This is seen as the ideal for open source AI.
  • Performance benchmarks show it's on par with O3 Mini Low, which is impressive for its size.
  • The presenter is keen to integrate it into coding tools.
• Cogito v1: A new company, DeepCogito, released Cogito v1 Preview, based on the Llama 3 70b model.
  • It's a 'hybrid reasoning' model – it can give direct answers or use special 'thinking tokens' for complex reasoning.
  • It's optimized for coding, science/tech tasks, following instructions, and has better multilingual and tool-using capabilities than similar-sized models.
  • Benchmarks show it significantly outperforms the base Llama 3 model and other reasoning models like DeepSeek R1 Distill 70b.
  • Why care? It's efficient enough to run locally on your own machine, which is great for privacy and control, boosting the open-source ecosystem. It also comes in smaller sizes.

Other AI News

• Midjourney V7: Midjourney released an alpha version of V7 for testing.
  • It apparently got less attention due to the hype around OpenAI's image generation.
  • The presenter notes some online disappointment, but competition is good.
  • Claimed improvements include better understanding of text/image prompts and higher image quality.
  • A key new feature is 'Draft Mode' – it generates images 10 times faster at half the cost. It's so fast the interface becomes almost conversational.
  • This speed contrasts sharply with the sometimes slow generation times of GPT-4 image models.
• Shopify's AI Mandate: Shopify CEO Tobi Lütke is pushing AI hard internally.
  • He believes AI will revolutionize what merchants can do on the platform.
  • He personally uses AI extensively and sees it as a massive shift in how work gets done.
  • His advice: learn AI by using AI extensively.
  • Crucial point: Using AI effectively is now a fundamental requirement for all Shopify employees. The presenter strongly echoes the idea that people using AI will replace those who don't.
  • AI must be used in prototyping, and AI usage will be part of performance reviews.
  • Teams need to prove they can't achieve goals with AI before asking for more staff or resources. The presenter thinks this is a smart move.
• Grok 3 API: Elon Musk's xAI released an API for Grok 3.
  • Grok 3 is considered a top-tier model, especially for coding.
  • Having an API means developers can integrate Grok into custom tools, agents, and workflows, which wasn't possible through the standard interface.
  • Pricing is detailed: Grok 3 Beta has a 131k token context window ($3/million input, $15/million output). Grok 3 Mini is much cheaper ($0.30/$0.50 per million tokens).
  • The price is seen as competitive for high-end closed models, but still way more expensive than open-source options.
• OpenAI & Jony Ive/Sam Altman Startup: There are reports OpenAI discussed buying the AI device startup co-founded by Apple's former design chief Jony Ive and OpenAI's own CEO, Sam Altman.
  • The presenter finds the potential acquisition structure a bit odd (CEO's side project bought by his main company).
  • Jony Ive was behind iconic Apple designs.
  • The goal is an 'AI-native' device. The presenter feels current options (Rabbit, Humane, Siri) fall short, and maybe Google's Gemini on Android is the closest we have.
  • This move would align with OpenAI's strategy: as pure AI smarts become common, focus shifts to personalization (Memory) and potentially dedicated hardware that deeply integrates this personalized AI, possibly replacing phones.
• Microsoft's 'Fast Follower' Strategy: Microsoft is intentionally letting others lead in cutting-edge AI development.
  • Their strategy, according to AI CEO Mustafa Suleiman, is to let pioneers spend big on R&D, then Microsoft follows 3-6 months later, learns from them, and builds potentially better or more integrated versions.
  • The presenter points out this isn't new for Microsoft (citing Teams vs. Slack, IE vs. Netscape).
  • This approach is seen as more cost-effective and part of Microsoft's effort to reduce its dependence on OpenAI, whose long-term vision might differ from Microsoft's.
  • Microsoft views becoming AI self-sufficient as 'mission critical'.