Ben Dickson

RT @bdtechtalks: The convergence of AI and neuroscience opens exciting possibilities for understanding human cognition and driving innovati…

I'm kind of bummed that Amazon cancelled "Wheel of Time" (they should have cancelled Rings of Power). I wonder if anyone can create season 4 with Veo 3. That would be a game-changer.

I use Gemini 2.5 Pro to cleanup the transcripts of my interviews and it amazes me every time how it can use the context of the interview and its internal knowledge to capture and correct very special nuances of the conversation that were missed by the ASR model.

This is... unsettling https://t.co/6vtJLFK9rL

LLM often struggle to figure out if they have enough context to answer a question or abstain. And using RAG can cause further confusion, as irrelevant information can put the model off. I spoke to @CyrusRashtchian on "sufficient context," a new technique to figure out whether

VentureBeat @VentureBeat

Why enterprise RAG systems fail: Google study introduces 'sufficient context' solution https://t.co/t29D3I5s0f https://t.co/LHvov9O0fy

Compared to Gemini 2.5 Pro <= 200k tokens (which is the limit of Claude models), Claude 4 models are 2.4–12X more expensive on input tokens and 6–30X more expensive on output tokens, even when running on Google Cloud. The architectural differences can't be that drastic to https://t.co/osgcFq1pmZ

The research and development on dLLMs is accelerating. Expecting great things in 2025...

AK @_akhaliq

dKV-Cache The Cache for Diffusion Language Models https://t.co/VusYoSLgTC

Llama 4 flopped Claude 4 underperformed 3.7 Google should take note and skip Gemini 4 in the future

Lisan al Gaib @scaling01

Claude 3.7 Sonnet better than Claude 4 Sonnet https://t.co/7Bfe9Ij3BI

Interesting study. Another aspect of the complexity of giving models knowledge and reasoning skills while also training them to know the limits of the their abilities.

Linxin Song @linxins2

🚨 We discovered a surprising side effect of Reinforcement Finetuning (RFT): it makes LLMs more confidently wrong on unanswerable questions. We call this the hallucination tax: a drop in refusal behavior that leads to overconfident hallucinations. 🧵 1/n https://t.co/mjcrp3HkMn

OpenAI is already burning through a lot of cash without making profits. And it spent $6.5 billion to acquire Ive’s firm. There is no way they’re going to do anything substantial without another huge funding round. https://t.co/hVpzuJgMia

New funding round incoming?

Sam Altman @sama

thrilled to be partnering with jony, imo the greatest designer in the world. excited to try to create a new generation of AI-powered computers. https://t.co/IPZBNrz1jQ

Feel the AGI?

John Carmack @ID_AA_Carmack

The demo we are showing is an RL agent on a laptop playing an unmodified Atari with a robotic controller moving the joystick and a camera looking at the screen. https://t.co/j36tH7lAJh

A friend had forgotten to update his payment details and his ChatGPT subscription expired. He asked me if he should renew. I told him to give the free version of Gemini a shot before renewing ChatGPT. He did not go back. I've seen this happen multiple times. Although the sample

Diffusion language models are moving into the mainstream. When Google does it, it means there will likely be gains at scale.

Google DeepMind @GoogleDeepMind

We’ve developed Gemini Diffusion: our state-of-the-art text diffusion model. Instead of predicting text directly, it learns to generate outputs by refining noise, step-by-step. This helps it excel at coding and math, where it can iterate over solutions quickly. #GoogleIO https://t.co/EmDAqNeXVv