Top stories in tech today: – Big Tech spends $226K a day to lobby Congress
– Meta and Microsoft are slashing thousands of jobs – Tesla is spending $25B to reinvent itself
– Instagram tests a stripped-down Snapchat rival
– Quick hits on other tech news
MACHINE LEARNING
-
Interesting Recent Open Source Projects: Mirofish, Paperclip, Hermes Agent, AutoResearch
By
–
-
Controlling AI output via prompts and verification for longer thinking
By
–
IME you still get a lot of control over this in the prompt itself. Even with the highest Pro settings, it thinks longer if you say to keep trying until some verification you propose passes etc.
-
Deep Learning: ConvNets and ViTs Training with Backpropagation
By
–
Deep learning. Largely ConvNets or ViTs trained with backprop
-
LLaTiSA: Difficulty-Stratified Time Series Reasoning Visual Perception
By
–
LLaTiSA
— AK (@_akhaliq) 24 avril 2026
Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics
paper: https://t.co/ODOruVboIu pic.twitter.com/ZM5G3vY3IKLLaTiSA Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics paper: https://
huggingface.co/papers/2604.17
295
… -
Deep Learning ConvNets: Core AI Beyond Language Models
By
–
The vast majority of them use deep learning, i.e. ConvNets trained with backprop. It's not LLMs but it is AI.
-
Deep Learning Without LLMs: Neural Net Backpropagation Focus
By
–
It's all based on neural net / deep learning trained with backprop.
But none of it uses LLMs, GPT, or other token-based generative architectures. -
LLMs vs Vision Systems: Neural Nets and Architecture Differences
By
–
Absolutely nothing. They both use neural nets and backprop. But LLMs are generative architectures trained on sequences of discrete symbols. Vision systems used in AEBS and other applications use ConvNets or ViTs trained to detect and classify from labelled samples.
-
Joint Embedding Architectures: Vision Encoders Beyond LLMs
By
–
Through 1. Vision encoders that are not LLMs. They are actually Joint Embedding Architectures that embed images and text description in the same space
2. Painfully exhaustive training on enormous amounts of declarative facts about the physical world. You can train them to answer -
Why LLMs fail at physical world understanding compared to JEPA
By
–
We still don't have AI systems that understand the physical world as well as a cat. JEPA are getting there. LLMs and other generative architectures are not. Generative architectures, particularly those that produce discrete tokens like LLMs simply ***DO NOT WORK*** for