I actually think the legacy completion API is superior to the newer chat creation API. The completion API takes all the input as a single string, whereas the chat API expects input as a list of messages. Given today's common use of agentic workflows, the model often has to process a large amount of input—RAG data, function calls, and other metadata—in addition to the user and assistant messages. The conversation itself might be just a small part of the overall prompt.
I’ve found it more intuitive to format the conversation as <conversation>...</conversation> within one unified text prompt, alongside all the other input data.
OpenAI’s new completion API supports both a single large string or a list of messages, which is great. I briefly experimented with the prompting tool in Claude—it essentially builds a big prompt that includes instructions and examples in XML format.
Great idea! The demo looks impressive. What are your thoughts on real-time translated captioning compared to AI voice? I guess it's still difficult to mimic nonverbal elements like laughter and pauses.
"synthesize large amounts of online information" does it heavily depend on the search engine performance and relevance of the search results? I don't see any mention of Google or Bing. Is this using their internal search engine then?
I actually like how NULLs behave in SQL. They mean "I don't know" In the modern programming language we all care about Null safety. But no matter how you model your data, you will always run into the situations when you don't know everything. So I believe NOT NULL is not very practical. NULLs in SQL handle these case very well - when the input is unknown your output is unknown
Those mini apps are built on the same web stack. I believe the main advantages of creating mini apps are that the platform provides identity (allowing you to know who the users are upon permission approval) and payment APIs.
I'm a heavy use of tmux integration in iterm2. this allows seamless mouse scroll in a tmux window. I haven't seen any other terminals that provide the same tmux support.
so the WebRTC helps with the unreliable network between the mobile clients and the server side. if the application is backend only, would it make sense to use WebRTC or should I go directly to realtime api?
i think cursor is the only player that forked vscode. i've been using cursor for a week and i really like it. it's able to auto complete/correct code in multiple places at the same time. other copilot extensions in vscode were not able to do this. but i assume microsoft will quickly add this feature in vscode and others will catch up soon. i don't know if it's a true advantage to stay as a fork of vscode.
I like the UI. The chat interface is a good fit for this task. How do you prevent or should you prevent users from entering "write me a fib(n) in python" in the chat? To me the chat is solely designed for table creation directives.
what do people think about installing new software from less well known companies and giving the app full access to all your files? how would i convince myself that this software won't upload my data in the background?
Cloudflare has a switch to block all the unknown bots other than the well behaved one. Would this be a simple solution to most of the sites? I wonder if the main concern here is that the sites don't want to waste bandwidth/compute for AI bots or they don't want their content to be used for training.
I'm curious here how do they make money? i heard that they have ai consulting service. is that sufficient to run the company? i don't think any researchers i know pay for hugging face.
ok does it mean an end to email? it's nearly free to send emails to anyone. for comparison, it's much more expensive to send linkedin messages or create ads on social networks. did anyone attempt to create a paid email service (pay to send)?
what are the pros and cons with pydantic models vs data classes? i like the runtime validation in pydantic models.
the pattern i landed on is that 1) only use primitive data types such as str, int, pydantic models and pandas dataframes to present data. 2) no classes with methods, no polymorphism, only use module functions.
i think llama.cpp has context caching with "--prompt-cache" but it will result in a very large cache file. i guess it's also very expensive for any inference api provider to support caching as they have to persist the file and load/unload it each time.
for any APIs related to money, should the currency be in strings as opposed to in floats? This will prevent accidental float arithmetic in the code. I always find it tricky to work with currency in javascript.
I’ve found it more intuitive to format the conversation as <conversation>...</conversation> within one unified text prompt, alongside all the other input data.
OpenAI’s new completion API supports both a single large string or a list of messages, which is great. I briefly experimented with the prompting tool in Claude—it essentially builds a big prompt that includes instructions and examples in XML format.