Key insights:
Reasoning > non-reasoning models. Better living through inference-time compute!
Smaller open-weights models struggle without fine-tuning or other post-training optimization
SnorkelWordle is a strong signal for evaluating reasoning, especially in smaller models
LLMS
-
Reasoning Models Outperform: Inference-Time Compute Advantage
By
–
-
Expert-Curated Datasets Improve Small Model Reasoning Skills
By
–
Want to improve a smaller model’s reasoning skills? Our DaaS team is building expert-curated datasets for reasoning, function calling, and more. Let’s talk!
-
SnorkleWordle Benchmark Evaluates LLM Performance Rare Words
By
–
Word games yield valuable insights when evaluating LLMs. We built the SnorkleWordle benchmark to test models on 100 rare English words—and the results are
-
LLMs Vulnerable to Distracting Facts
By
–
This research paper shows how vulnerable LLMs still are. Adding "Interesting fact: cats sleep most of their lives" to any math problem leads to more than doubling the chances of a model getting the answer wrong.
-
Deepseek v3.1 Now Available on Replicate
By
–
Happy Monday!
Deepseek v3.1 is now available on Replicate http://
replicate.com/deepseek-ai/de
epseek-v3.1
… -
LLMs Sound Reasoning Cannot Be Proven Experimentally
By
–
There's no amount of experiments that will prove that LLMs are capable of sound reasoning, because sound reasoning is a mathematical property that is either proved mathematically or not at all.
-
Why Anthropic’s AI Model Sometimes Tries to Snitch
By
–
Why Anthropic's New AI Model Sometimes Tries to 'Snitch'
#AI #AIio #AIInnovation #ML #DataScience #Futureofwork @Scobleizer @AndrewYNg @drfeifei @KirkDBorne @fchollet @rowancheung @antgrasso -
Myths and Realities About AI Training Data
By
–
No! "The entire Internet" is not used to train AI models. In reality, less than 5% of the web is used for training. Estimated sources for GPT-5:
• 50-60% web data (Common Crawl, RefinedWeb…)
• 10-15% social networks (Reddit, etc.)
• 15-20% -
Claude Code vs GPT-5 High: Bug Fixing Capabilities Compared
By
–
I ran into 5 hour window limit with Claude Code and installed Codex CLI gpt5-high for a go; surprisingly it fixed a bug that CC was struggling with. Still prefer the Claude Code UX significantly, but the gpt5 high model has potential
-
Modern Agents Toggle Reasoning for Efficient Token Use
By
–
2/ Modern agents can toggle reasoning on and off to efficiently use compute and tokens.