🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

OpenAI: GPT-4.5 is our most creative model yet. Me: Doubt.

Don't be disappointed that GPT-4.5 isn't smarter than o1. Scaling up pretraining improves responses across the board. Scaling up reasoning improves responses a lot if they benefit from thinking time and not much otherwise. Wait to find out how the improvements stack together.

Flattering the model makes a comeback.

Looks like gpt-4.5-preview is 25% faster than when I tried it yesterday, although still many multiples slower than 4o.

Even the newly released Sesame knows the atoms joke. sesame.com/research/cross…

Today I’m excited to announce our Series A fundraise by Norwest 🔥 Agents have the potential to automate the majority of knowledge work - whether it’s financial due diligence, support resolution, PRD generation, contract review. Building these agents requires both data and

`npm install -g @anthropic-ai/claude-code` there's no more waitlist. have fun!

If you're working on voice AI, watch this space.

Mistral OCR is nice and fast but other models outperform it on document processing. We did a comprehensive benchmark on Mistral OCR and compared it against a comprehensive set of different LLM/LVM-powered parsing techniques - these direct parsing using gemini

Every programmer I interviewed was allowed to use ChatGPT/Cursor/Copilot. I expect that to be the norm in a few years. At the present moment it's relatively straightforward to tell the difference between someone who actually understands the code and someone who doesn't. That

leehanchung.github.io/blogs/2025/03/…

From Yann LeCun on greencard/visa revocations and deportations of scientists.

Packaging a library for NPM that works on multiple runtimes is way more challenging than it should be. The Gemini team seems to have taken one of the cleaner approaches I've seen to date. Will need to give it a shot in a future project.

Alex Albert Anthropic Am I right in understanding that there's no additional execution of any separate code here at all? You tell Claude "put your thoughts in the think tool if you need to" but it's effectively a null-op - it's effectively a prompting hack that encourages Claude to "think out loud"

One of the reasons why LLM hallucinations are so hard to deal with is that the models generally output responses in a confident tone of voice. AI "gaslighting" or evaluating underlying model certainties may be a way to uncover actual confidence vs. good bluffs.

The "inputs only" reinforcement learning mechanism from Databricks is an interesting idea. databricks.com/blog/tao-using… Looks like the crux of it is a custom-developed proprietery reward model, and not LLM as a judge though.

This Claude interpretability blog is one of the most interesting ones I've read so far this year. We all know what LLMs output, but how do they choose their output? With billions of parameters it might seem impossible. But the result is human-like? anthropic.com/research/traci…

When chess computers first started getting good, Anand and others pushed for a variant where humans and computers would work together on a team. That variant gradually grew out of favor because the computers got so good that the humans weren't adding anything. We're not ready.

Interesting Chinese stocks have been some of the least affected. From the NYT

Due diligence: When you find out someone is dumber than bricks before you spend 100 million putting them into power.