Which AI have you used recently to believe that AI still codes like crap in 2026? Because there have been some serious advances and its vision is not far-fetched, I think.
MACHINE LEARNING
-
Cold weather prompts GPU usage to fine-tune large XGBoost models
By
–
It's getting pretty cold over here. Time to fire up all the GPUs and fine tune some big XGBoost models. @trainxgb
-
LeCun: child perceives more visual data than LLM
By
–
This is false: LeCun does the calculation in a video, I need to find it: a child in their first 4 years of life perceives more visual data than the largest LLM trainings use.
-
Reflecting on the black-box nature of neural network weights
By
–
Yeah, I've been reading through the source code too. It's quite interesting. What's really interesting is that a lot of it is run by neural networks that we can't see inside of; we don't know the weights, for instance, but we can see the effect. My feed is generally better today
-
Databricks Co-founders Speaking at Scaled ML 2026 Conference
By
–
Databricks co-founders @matei_zaharia and @istoica05 will be speaking at #ScaledML 2026! This event brings together the creators behind systems like Apache Spark™, CUDA, TensorFlow, ImageNet, Tesla FSD, and more to explore large-scale learning, distributed systems, and next-generation AI hardware. Join us on January 29th: scaledml.org/
→ View original post on X — @reza_zadeh, 2026-01-20 20:10 UTC
-
Gemini 3 Pro A/B Tested on AI Studio with Robot SVG Bench
By
–
A true Gemini 3 Pro (GA?) is being A/B tested on AI Studio. Robot SVG bench
-
DeepSeek repo references new model1 and FlashMLA attention kernels
By
–
One of the DeepSeek repositories got updated with a reference to a new “model1” model. “FlashMLA is DeepSeek's library of optimized attention kernels, powering the DeepSeek-V3 and DeepSeek-V3.2-Exp models.” Soon?
-
Ingenium Group 2025 Annual Report: Publications, Citations, and Innovation
By
–
Ingenium Group 2025 Annual Report released! – 37 publications, 7 Q1 journals – 2,000+ citations in 2025 – Launched Ingeniot (IoT + AI lab) & global Ingenius Think-Tank – Leading EU projects in renewable energies Proud of our team’s impact on innovation & sustainability.
→ View original post on X — @faustospain, 2026-01-20 10:01 UTC
-
Transparency and Weight Accessibility in Open Source AI Models
By
–
It's a good thing that it's open source so that we can analyze it. Whether the algorithm is any good or not relies on the neural network that is being called by the structure we can see in the source. However, since we can't see the weights, it's not completely "open source."
-
Simpler prompting techniques derived from research papers
By
–
This is simpler than Lyra but incorporates the actual research-backed techniques: ∙ Clarifying questions (ICLR 2025)
∙ Constraint specification (all major guides)
∙ Output format control (Anthropic, OpenAI)
∙ Iterative refinement (feedback loop)