Prediction: unless patched, GPT's tendency to say delve will be assimilated into North American style English. Soon it won't stand out as odd anymore. More broadly, RLHF'ed LLMs will shape cultural norms in unexpected ways.

This is something I've worked on for a while! You can save the activations of one LLM call and reuse them for a follow-up that overlaps with the first. This means asking a question about a big codebase can take 30 seconds the first time and 1s after that!

prediction: when all is said and done Elon Musk's $42B purchase of Twitter of Twitter will be seen as a bargain. Not because of the platforms economic success, but its outsize sociopolitical influence.

12-hour timelapse of American Airlines, Delta, and United plane traffic after what was likely the biggest IT outage in history forced a nationwide ground stop of the three airlines.

Here’s a video of his brother admitting it

Phenomenal response, from the founder of Deepseek, on moats

Bearish on Cursor (closed source, send your code to their servers, forced subscription) Bullish on Zed (open source, fast, private, existing anthropic partnership, supports API keys)

1/6: A recent paper shows that that LLMs are "self aware": when trained to exhibit a behavior like "risk taking", LLMs self report being risky. In a recent blog post, we explore what's happening here: some self awareness behaviors are caused by a simple learned steering vector!🧵

One failure mode that I’ve repeatedly fallen into is thinking “surely someone smarter than me must have already figured this out” when in fact no one has.

Our new Google DeepMind paper, "Lessons from Defending Gemini Against Indirect Prompt Injections," details our framework for evaluating and improving robustness to prompt injection attacks.

even if model "neutrality" were mean perfectly representing the views & biases of the population in the model, the result would mean cementing the status quo.

Officially validated IMO gold medal, purely via search in token space, achieved in 4.5 hrs (unclear at what compute cost). The solutions read nicely as well deepmind.google/discover/blog/…

correct me if im wrong but it seems like: - the theme of the Dan Wang book, and the general elite consensus now is that “industrial process” is a technology that lives in the heads of people and that it was a mistake to let so much “low value” industry be offshored due to the

This chart is potentially misleading. It compares the latest sonnet model to older models (from March/April). Also, attacks in this data set were optimized against those other models, but not against sonnet 4.5 (!). It would likely do worse against tailored attacks.