Finished testing Claude Opus 4 and Claude Sonnet 4 on my personal eval set. I am VERY impressed. Claude Opus 4 absolutely dominated other models in both coding and writing tasks. It is the best performing model for all 4 tasks given. It is worth nothing this model is very

I have always believed that instruction-tuned models are the same type of models as reasoning models in how they work fundamentally (predicting the next token). But more recently I'm exposed to another group of people who think that by embedding thinking directly into the model.

Screenshot from ccusage for June 2025.

Damn it. Looks I lost my ability to send emails to my customers on Mailchimp. Any good alternative for just over 500 emails?

I just realized that people don't really want the truth. They just want confirmation that they are right. Gonna to try apply this principle to marketing. .

How to spot AI-generated content (July 2025 edition):. - "It is not just . , it is . ".- "Not gonna lie, . ".- ". The bottomline is . ".- "Key takeaway: . ".

I only managed to decipher the phrase "[155]love@pliny[157]" with the help of Claude Code. What's the actual trick here?

The truth is, we don't need so many models (as users and AI app developers). People have the patience and incentive to maybe compare two or three top models, and that's it. They don't care about the 100 other models that's marginally better on some niche areas. For me I've.

It's actually quite fun to see AI skeptics send you links of AI-generated clickbaits that shows various AI fails, which are either outdated news, or cases of people who don't know how to use AI.

Wow. Luckin Coffee & @duolingo collab spotted in Singapore!. As a daily user of both, I feel blessed!

Link to the GitHub source. You can Google translate: .

Detailed leak on how Huawei Panguo models have been trained.

Compulsory reading if you want to understand how this mental model works:

Because everything is probabilistic, you are better off writing tokens that have a higher chance of activating the correct features to output the desired tokens.

I belive that the right way of writing good prompt is to think about features, nodes and activations, not how you would write an essay or instruction manual.

It would probably take me a few hours of research to find this method and implement it correctly, but it only took Claude Code less than 5 minutes.

Is this AGI? Claude Code just found a trick to get X content via X embed API, bypassing normal bot filters. To be fair, this method can be found on GitHub, but still impressive.

I wrote a blog post debunking the most popular prompt for vibe check, 9.9 vs 9.11. The truth is, the answer depends on the context:.

@16xEval Compare versions by viewing the changes visually

@16xEval Firing the same prompt and context to 3 different top models, and then pick the best output (view visual diff and assign human rating) is really effective.

When it comes to writing, there is no single model that dominates the other. Depending on the task, the context and lots of other random factors, one model would give you better result than the other. I have observed this while building and using @16xEval. The best workflow I