15 posts found
Jensen Huang said AGI has arrived. OpenAI's own president hinted at the same thing. But the benchmark propping up that celebration hides an uncomfortable fine print. I explain why the engineering truth behind Astra is far less magical than the pitch.
A white paper bringing together researchers from Google, Stanford, Harvard, and Princeton proposes something that caught me off guard: maybe AGI won't come from text, but from video models like Veo 3. I explain why that hypothesis isn't as crazy as it sounds.
WeatherNext, from Google DeepMind, gave hurricane forecasting a full extra day of lead time — and helped anticipate the brutal intensification of Hurricane Melissa in Jamaica. I explain why this is one of the AI uses that impressed me most this year.
9 models. 145 political questions. BaiChuan refused 60%. DeepSeek refused 36%. GPT-4o refused 0%. PNAS Nexus published the first comparative analysis of LLM censorship — and proved that when the government approves the model before launch, the AI goes silent.
A paper accepted at ACL 2026 proved: swapping fancy vocabulary for simple words boosts GPT-4o, DeepSeek and LLaMA accuracy by 8%. Adam's Law explains why — and frees you from the 'corporate speak' sabotaging your responses.
A math professor sent an absurd paper to an obstetrics journal. Accepted in days. NeurIPS 2025 published 100+ hallucinated citations across 51 papers. 1 in 277 PubMed articles in 2026 has fabricated references. And this data will train the next generation of AI.
GPT-4o: from 91% with 5 words to 15% with 40. In mixed conditions: 1%. GPT-5, Claude Opus 4.1 and Gemini 2.5 also failed. A PNAS Nexus paper proved: transformers have no 'referee' for resolving attention conflicts. It's an architectural limitation.
Nature Medicine, June 12, 2026: GPT-5.2, Claude Opus 4.6 and Gemini 3.1 Pro beat OpenEvidence and UpToDate on ALL tests. 100 real physician queries. 12 clinicians evaluating blind. 1,800 annotations. And the 'specialized' tools didn't even beat Google Search AI Overview.
Cosmos 3 launched June 1, 2026: the first open omnimodal model that understands and generates text, images, video, audio AND robotic actions in one architecture. Jensen Huang: 'The big bang of physical AI is coming.' And he opened the entire code.
A Journal of Marketing study with 7 experiments and data from 27 countries proved: the more you know about AI, the less receptive you are. The mechanism? 'Magical awe.' Demystifying AI reduces adoption. And that's a massive problem.
+8.3% accuracy. Up to 2.4x faster. 75.6% fewer tokens. Training cost: $4.27. RecursiveMAS proves that when AIs stop 'talking' and start 'thinking together,' everything changes.
AlphaFold solved protein shape. But shape is just a photo. MIT's VibeGen designs molecules by their movement — creating proteins that never existed in nature. Literally: vibe coding for biology.
Stanford proved that changing only the harness — without touching the model — creates up to 6x performance swings. Meta-Harness ranked #1 on TerminalBench-2, beating all human solutions. The problem was never the model.
Yann LeCun left Meta, raised $1 billion and said LLMs are a 'dead end.' I thought it was hyperbole. After deep research, I changed my mind. Partially.
What if the specialist you're looking for inside the AI is already there — just hidden? MIT's Neural Thickets concept made me rethink everything I thought I knew about fine-tuning.