Hacker News Reader

Comment by jasonjmcghee | Hacker News Reader

Surprised more comments aren't mentioning deepseek has this feature (for free) already. Assuming this is why OpenAI scrambled to release it.

The examples they have on the page work well on chat.deepseek.com with r1 and search options both enabled.

Do I blindly trust the accuracy of either though? Absolutely not. I'm pretty concerned about these models falling into gaming SEO and finding inaccurate facts and presenting them as fact. (How easy is it to fool / prompt inject these models?)

But has utility if held right.

View on HN

Not really accurate. The "Search" functionality you're describing in DeepSeek is comparable to OpenAI's existing "Search GPT." OpenAI's recent announcement refers to a more advanced capability, similar to Gemini's existing "deep research" feature. DeepSeek's current offerings are significantly more limited in scope.

Doesn't seem like access is available to try "deep research" yet on OpenAI, so I can only speak to what I tried, which was their examples on the blog post (using DeepSeek w/ R1 + Search) and results were pretty similar.

AFAIK OpenAI's current offering uses 4o, and it does a web search and then pipes it into 4o. I'm guessing adding CoT + other R1/o3 like stuff is one of the key effective differences. But time will tell how different it is. Maybe it's a dramatic improvement.

SearchGPT is bad because its underlying model is not a reasoning one. Deepseek one mentioned above is closer to deep research than searchgpt.

Are you unaware that there is a "Deepthink (R1)" button right next to the "Search" button on DeepSeek's Chat app. Its been there for some time, even before all the hype regarding R1.

I'm well aware of that. That is not what openai calls "deep research".

I wish Kagi would work with similar performance. Their lenses feature is perfect for this and they already filter out most of the SEO spam based on trackers and other typical red flags.

Reply on news.ycombinator.com