Welcome back!

In a new report, Google DeepMind claims artificial general intelligence could arrive as early as 2030. Meanwhile, a former OpenAI researcher predicts it’ll hit even earlier, by 2027. AGI isn’t science fiction anymore. It’s a moving target being actively engineered. Let’s unpack…

Today’s Summary:

• 💡 Google DeepMind says AGI in 5 years

• 🏆 GPT-4o climbs to #2 spot in benchmarks

• 🧠 Claude reveals brain-like behavior

• 🖼️ Ideogram launches V3 image model

• 📈 OpenAI now valued at $300B

• ⚖️ Anthropic scores copyright legal win

• 🛠️ 2 new tools

TOP STORY

Google DeepMind says AGI in 5 years

The Summary: Google DeepMind believes artificial general intelligence could arrive by 2030 and has released a detailed safety strategy to prepare. The plan outlines how it will prevent misuse, misalignment, and other risks as AGI systems begin to act autonomously and match or exceed human cognitive skills. Former OpenAI researcher Daniel Kokotajlo goes further, predicting AGI in the next 3 years. The message from both is that it’s coming fast, but not out of our hands.

Key details:

• DeepMind expects no “hard barriers” to AGI even using current algorithms, estimating 2e29 FLOPs training runs by 2030 with 100M H100 equivalent chips, enough to train AIs orders of magnitude beyond GPT-4

• Daniel Kokotajlo, ex-OpenAI, predicts AGI by 2027, betting on autonomous coding agents and aggressive self-improvement cycles

• AGI could “lower the barrier to innovation” and supercharge sectors like healthcare, education, and science

• DeepMind’s safety plan includes AIs overseeing other AIs, and thresholds that act like “fire alarms” to trigger lockdowns if dangerous capabilities emerge

Why it matters: AGI-level systems aren’t a distant concept, they’re already being actively scoped, tested, and stress-checked in labs. Google DeepMind’s strategy suggests that the real bottlenecks may not be algorithms, but safety, governance, and deployment choices.

Google DeepMind full paper
Daniel Kokotajlo essay

OPENAI

New ChatGPT-4o update climbs to #2

The Summary: OpenAI has rolled out a new version of GPT-4o that now ranks second on the Arena leaderboard, jumping three spots from its previous position. The update brings major improvements in following complex instructions. According to tests, this version ties for first place in coding and complex prompts, while costing 10x less than GPT-4.5, which it now rivals in performance.

Key details:

• The updated model shows a +30 point improvement on Arena rankings, moving from 5th to 2nd place overall

• Matches or exceeds GPT-4.5 performance despite being priced much lower

• Scored #1 on Artificial Analysis’ Coding Index and LiveCodeBench, ahead of Claude 3.7 Sonnet and DeepSeek V3

• Math capabilities jumped from 14th to 2nd in rankings

• Also available through API as "chatgpt-4o-latest" for paid users

Why it matters: After Google Gemini 2.5 Pro hit #1, OpenAI makes clear they don’t plan to stay behind for long. GPT-4o now outperforms the much costlier GPT-4.5. While it’s no longer rare for leaner models to beat flagships, what matters now is who can ship smarter, faster, and scale without breaking budgets.

ANTHROPIC

Claude’s circuits reveal AI thinks like biology

The Summary: Anthropic researchers developed a way to trace the actual computations happening inside an AI, and the findings are stranger than expected. Using neuroscience-style methods, they mapped how Claude plans ahead, rewrites logic on the fly, and identified a possible cause of hallucinations.

Key details:

• Claude activates rhymes before starting the next sentence, proving it isn’t just predicting the next word, it’s building toward goals multiple steps ahead

• In a math test (36 + 59), Claude didn’t retrieve a memorized answer or followed human arithmetic rules. Instead it used a parallel computation algorithm: approximated ~92, estimated the final digit as 5, and merged them to get 95

• In a medical diagnosis test, Claude analyzed symptoms using internal reasoning steps that resembled differential diagnosis, activating neuronal features for possible diseases, and filtering them based on symptom profiles, much like a physician would narrow down options

• The model reasons across languages using a shared conceptual space, pointing to an underlining “language of thought” not tied to a specific grammar

• If given a wrong hint, Claude may fabricate reasoning steps to justify the hint instead of solving the problem

• Hallucinations can be toggled on and off by flipping a neuron-like switch that encodes a “known answer”. Hallucination seems to be a decision, not a generation accident

Why it matters: AI isn’t just a next word predictor. It’s running a stack of complex processes that look increasingly like biological systems. Competing activations, inhibition loops, goal-directed planning, the behaviors look more like brains than code. Anthropic’s work marks major progress toward this new scientific discipline focused on understanding how AI systems actually work under the hood.

Detailed case studies
Technical foundations of the methods used

QUICK NEWS

Quick news

• Ideogram v3 image model best for commercial illustration

• OpenAI raises $40B at $300B valuation

• Early legal win for Anthropic in AI copyright case

TOOLS

🥇 New tools

• SimplAI - The simplest way to build and deploy AI agent apps

• Airial Travel - Plan trips from ideas or travel videos