TechTalks

TechTalks

Home
Archive
About
Inside Nvidia's new technique to optimize long-context inference and continual learning
By treating language modeling as a continual learning problem, the TTT-E2E architecture achieves the accuracy of full-attention Transformers on 128k…
 • Ben Dickson
Meta’s new VL-JEPA model shifts from generating tokens to predicting concepts
Meta’s VL-JEPA outperforms massive vision-language models on world modeling tasks by learning to predict "thought vectors" instead of text tokens.
 • Ben Dickson
How reinforcement learning changed LLM tool-use
A look at the evolution of LLM tool-use, from supervised fine-tuning to Reinforcement Learning (RLVR) and agentic applications in large and specialized…
 • Ben Dickson
Inside URM, the architecture beating standard Transformers on reasoning tasks
The key to solving complex reasoning isn't stacking more transformer layers, but refining the "thought process" through efficient recurrent loops.
 • Ben Dickson
What (I think) makes Gemini 3 Flash so good and fast
Google didn’t reveal a lot of information about its new Flash model. So we had to speculate a lot on what is going on under the hood.
 • Ben Dickson
Nvidia sets a new bar for open source models with Nemotron 3
As the industry shifts from chatbots to multi-agent workflows, Nvidia's Nemotron 3 offers a blueprint for efficient, long-context reasoning.
 • Ben Dickson
AI benchmarks are confusing. Here's why.
AI labs are racing to overtake each other on key industry benchmarks. But this intense race has stripped the benchmarks of most of their value.
 • Ben Dickson
Salesforce takes a new approach to web agents with WALT
The framework abstracts away the chaos of dynamic layouts, allowing AI to focus on high-level planning and tool-use instead of low-level clicks.
 • Ben Dickson
Poetiq crushed ARC-AGI-2 at half the cost of Gemini 3 Deep Think. Here's how.
The verified solution achieves 54% accuracy on the semi-private test set, outperforming Gemini 3 Deep Think at less than half the cost.
 • Ben Dickson
The magic sauce that makes DeepSeek-V3.2 so damn efficient
DeepSeek-V3.2 is a top-5 LLM, sitting next to the likes of Grok 4 and GPT-5. But what is more impressive is its cost-efficiency and magnificent…
 • Ben Dickson
What OpenAI's 'Code Red' says about the current state of the AI race
OpenAI’s problem is not that it doesn't have the best model anymore but that the general feeling is that it has fallen behind.
 • Ben Dickson
What comes for LLMs after reinforcement learning with verifiable rewards (RLVR)?
Reinforcement learning from verifiable rewards (RLVR) ushered in a new generation of reasoning models. Now, researchers are looking beyond RLVR for the…
 • Ben Dickson
TechTalks
TechTalks
In-depth discussions about machine learning, deep learning, reinforcement learning, neural networks, artificial general intelligence, AI business, and other technology trends.
Recommendations
View all 12
AI Tidbits
AI Tidbits
Sahar Mor
The VC Corner
The VC Corner
Ruben Dominguez
Colligo
Colligo
Erik J Larson
AI Disruption
AI Disruption
Meng Li
The Founders Corner®
The Founders Corner®
Ruben Dominguez
© 2026 Ben Dickson · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture