First, I have watched the free improvement of frontier models surpass the gains from retraining, many times now. Squeezing more out of the models that already exist, or simply doing nothing and waiting, is a real strategy and it often pays better. The fair comparison is not against today's frontier but against whatever ships while you are still maintaining your fine-tune.

Second, the $500 training bill is the cheapest line item in this story. The expensive parts are creating the data and maintaining the model afterwards. How many use cases can actually produce 177k scored episodes? Here they had to generate them synthetically from Amazon Berkeley Objects. To me, that dataset is the strongest evidence in the article of how hard fine-tuning is to apply: if the data existed naturally, nobody would need to manufacture it.

Comments URL: https://news.ycombinator.com/item?id=47923778

Points: 1

# Comments: 0

I'm curious, is anyone using mypy in a production environment? What are your thoughts on introducing mypy from the beginning of development? Are you for or against it?

Comments URL: https://news.ycombinator.com/item?id=36518162

Points: 2

# Comments: 1

I agree that abstraction can be leaky, and it's true that you often end up writing some SQL-like code.

However, isn't writing raw SQL potentially dangerous, considering aspects like security, access control etc?

I want to learn your thoughts!

I also liked Heroku, so it's sad to see their status now.

Also, I live in Japan and can perhaps provide an example:

In Japan:

- Public infrastructure investment reached its peak in the mid-1990s.

- The birth rate peaked in the mid-1970s.

These tools likely don't permit users to directly manipulate an actual database table. Therefore, I believe they adopt alternative approaches.

My hypothesis is that they might utilize NoSQL databases and treat each row as a discrete item.

However, ensuring reference consistency without using real table features might be very challenging. Some tools offer a "calculated column" feature, which stores the results of complex functions that take the results of referenced tables as input. Caching and data change notifications might do the job, but is it really feasible?

Furthermore, I suspect that performance might be severely degraded in such an architecture if the number of rows grows significantly.

Any thoughts?

Comments URL: https://news.ycombinator.com/item?id=36475866

Points: 1

# Comments: 2

I know there are plenty of great projects for terminal AI assistants already out there. But, in my experience, none of these tools completely meet the criteria I consider essential:

Terminal Reading: Most tools are unable to read your terminal, which means you have to manually copy and paste the terminal output elsewhere.

Simple API: Who wants to remember a plethora of complex commands and options? This is the very reason I use these tools in the first place.

Multi-turn Capability: Most tools are designed for single-turn queries due to their lack of memory, significantly limiting their utility.

Direct Command Execution on zsh/bash: Some tools execute commands in a wrapped environment or REPLs to capture the terminal output. This approach can break some commands, which I find undesirable.

Personally, I use this tool very conveniently. I open chat.openai.com less often because I can get enough answers from GPT-3.5 or GPT-4 without leaving the terminal!

Any feedback or comment is appreciated!

Comments URL: https://news.ycombinator.com/item?id=36289298

Points: 1

# Comments: 1

I hope open-sourced models will achieve better performance soon.

Stacked neural networks (deep learning) / CNN / Transformer

From the application perspective, there are also areas that have seen significant progress:

AlphaGo (An application of reinforcement learning) / AlphaFold (An application for biological sequences) / ChatGPT (An application of Large Language Models / Reinforcement Learning from Human Feedback)

You may pick different ideas. However, what I want to say is that these great advancements happened approximately once every two or three years. It may take some time before we witness another major leap.

In truth, the majority of research and development does not contribute to exponential progress. It's sad. However, this does not render such effort meaningless. Although it may seem as though there's too much competition within a narrow field, this may not necessarily be detrimental to humanity.