Click any tag below to further narrow down your results
Links
This DeepMind essay argues that smarter AI alone won't reshape the economy much; the big shift comes when AI agents can coordinate with each other the way effective human teams do. It maps capability thresholds for competence and coordination and tracks how market structure, employment, and wages would change across them.
- Most AI benchmarks test individual agents in isolation, but economic outcomes come from organizations, where a well-structured team can beat a stronger solo performer.
- In a 100-person quiz across six domains, people were jagged: skill in one area barely predicted skill in another. Pairing complementary specialists outperformed random pairs by 32%.
- AI agents can be copied on demand at the cost of compute, so once a model matches human capability, human wages become tied to compute prices. Models also tend to be jagged in similar ways, so adding copies doesn't add breadth the way adding a complementary human does.
Google is releasing Gemini 4 Carbon, a frontier model it says leads on coding, agentic work, and cybersecurity tasks like finding, exploiting, and patching vulnerabilities. Access is rolling out first to a small group of cyber defenders and trusted partners before a broader release.
- Carbon scores 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching, which Google describes as state-of-the-art.
- On internal tests it cuts unsupported claims by about 2x versus Gemini 4 Argon, and it resisted every prompt-injection attempt in Gray Swan's benchmark, with a 0.0% attack success rate.
- Rollout is gated: it goes to selected cyber defenders first, then trusted partners through the Fairwind Program, with general developer and consumer access to follow once guardrails are refined.
A Google and MIT FutureTech study finds scientists save about 7 hours a week using AI, but nearly half now report a growing backlog of untested hypotheses. The piece, an interview with study co-lead Mihai Codreanu, looks at why faster idea generation isn't yet producing faster discoveries.
- The study drew on 15M Gemini interactions, 2,690 specialized models, and a survey of 637 US and UK researchers; about half of respondents said their bottleneck had shifted to downstream work like experimentation and data collection.
- A "verification tax": 88% of scientists spent more than 10% of their saved time checking AI outputs, and nearly half spent over a quarter of it.
- AI may nudge researchers toward safer, data-rich questions (a "streetlight effect"), even as it widens their access to ideas from other fields.
@Dipanshu_AI on X
The post promotes a 60-minute Cambridge lecture by Google DeepMind CEO Demis Hassabis on where AI is heading. The author's pitch is that an hour of this talk beats an hour of streaming TV for understanding AI's near-term future.
- The post frames the lecture as a substitute for an hour of Netflix, pitching it as a high-value use of evening time.
- It claims the talk will teach more about the future of AI than most people will pick up over the next five years.
- The author urges readers to bookmark it and watch in full, with no specifics on the content beyond the speaker and venue.
This article introduces the Gemma 4 family of models from Google DeepMind, detailing their architectures and improvements over the previous version, Gemma 3. It highlights key features such as interleaved attention layers and efficiency enhancements in global attention mechanisms.
- Gemma 4 comes in four sizes—E2B, E4B, 31B, and a 26B MoE model that only activates 4B parameters at inference—and handles text, images, and audio
- Local-to-global attention layers are interleaved at a fixed 4:1 ratio in smaller models and 5:1 in larger ones, with sliding windows of 512 tokens (E2B/E4B) or 1024 tokens (larger models) to cut compute costs
- Global attention memory use is reduced through Grouped Query Attention, a K=V trick that makes keys and values identical, and p-RoPE for improved positional encoding