🔥 Altman Says AGI by 2025

PLUS: Qwen AI Coder Runs on Your Laptop

Welcome back!

AI development is shifting gears. Traditional training scaling methods seem to be losing steam, but AI capabilities are still advancing. Companies like OpenAI and Anthropic are pivoting from brute-force scaling to more nuanced engineering — suggesting that smarter techniques, not just bigger models, are the way forward. Meanwhile, Sam Altman says AGI could arrive by 2025. Let's unpack...

Today’s Summary:

  • 💡 The path to smarter AI models is pivoting

  • 🥇 Qwen breaks records with 32B open-source AI coder

  • 🎥 Google Vids presentation tool for the workplace

  • 🧬 Google AlphaFold3 now open-source

  • 🛡️ Anthropic partners with Palantir to sell AI to defense customers

  • 🐝 Meta's nuclear-powered AI data center halted by bees

  • 🛠️ 2 new tools

TOP STORY

Next-gen AI models will break from conventional scaling laws

The Summary: While Sam Altman says AGI could arrive by 2025, there are signs that traditional AI training methods may be facing scaling barriers. This contrast between public optimism and technical challenges behind the scenes reveals the complexity of AI’s progress. OpenAI's upcoming model “Orion” has been delayed and is rumored to bring nuanced improvements, similar to delays with Anthropic’s Claude Opus 3.5. But both companies are now pivoting beyond brute-force scaling by working on deeper technical challenges, maintaining optimism for rapid advances in AI capabilities.

Key details:

  • OpenAI “Orion” is rumored to bring less dramatic improvements than previous generations

  • Anthropic CEO reveals massive leap in compute requirements for foundation models - from $100M for GPT-4 to $100B by 2025

  • Quality training data is becoming scarce; research firm Epoch AI predicts exhaustion of usable text data by 2028

  • OpenAI is focusing on innovative inference techniques with the o1 model series, reaching PhD-level performance. The “full” o1 model is anticipated as the next major release

  • Google appears ready to launch Gemini 2.0 soon

  • AGI timeline: Altman predicts 2025 arrival, while Amodei projects 2026/2027

Why it matters: These developments indicate a change in how AI technology advances, moving from brute force scaling to more refined engineering techniques, while AI capabilities are expected to keep progressing. The industry's response to these challenges will be key to achieving “powerful AI” in the coming years.

“If you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there [powerful AI] by 2026 or 2027. We are rapidly running out of truly convincing blockers, truly compelling reasons why this will not happen in the next few years."

Dario Amodei, Anthropic CEO
AI CODING

Qwen breaks records with free 32B AI coding assistant

The Summary: Alibaba's Qwen team released Qwen2.5-Coder, a free open-source AI that matches GPT-4o coding abilities. The new model works on laptops - running on M4 MacBooks with surprising efficiency. Early users report the 32B model needs just 18GB RAM, making it usable on consumer hardware. Qwen2.5-Coder speaks 92 programming languages and scored 92.7% on HumanEval, setting new records for open models.

Key details:

  • First open model trained on 24 trillion tokens, breaking previous records

  • Outperforms GPT-4o on several benchmarks including CrossCodeEval and RepoEval

  • Runs at 28 tokens/second on a single RTX 3090 GPU with 32K context

  • Mac M1 users reported 9 tokens/sec speed in LM Studio with 18GB RAM usage, and MacBook Pro M4 Max users reported 23 tokens/sec with 22GB RAM usage

Why it matters: Enterprise-grade AI coding tools are moving from costly cloud services to accessible local options. Qwen's efficiency on consumer laptops demonstrates that cutting-edge AI capabilities no longer require massive computing resources.

TOOLS

Google launches Vids AI for the workplace

The Summary: Google has introduced Vids, a browser-based video presentation app that generates storyboards, scripts, and voiceovers from documents and prompts. Teams can use it to create help videos, training content, and company updates. The rollout started across Google Workspace tiers, with AI features free until 2026.

Key details:

  • Vids generates complete video drafts from Google Drive documents

  • Includes a built-in recording studio with screen capture, selfie video, and audio recording with teleprompter

  • Free access to AI features until December 31, 2025

  • Real-time collaboration lets team members work on videos together

Why it matters: With visual communication becoming essential in the workplace, Vids positions Google to capture the growing market for internal corporate video content, challenging traditional presentation tools.

QUICK NEWS

Quick news

TOOLS

🥇 New tools

  • Voistory - Your journal with no writing

  • Sona - Conversations in, insights out

That’s all for today!

If you liked the newsletter, share it with your friends and colleagues by sending them this link: https://thesummary.ai/