This resonates with my experience. At Morph we use gemini for well specified point coding tasks, and it does very well across millions of lines of code every day. We also use claude code as an engineering tool for our own codebase and it does better at being adaptive and for working on open ended issues.
It would be nice to see a side-by-side comparison with Github on pricing and features. We are using github and creating hundreds of repos everyday without any issues (except for the occassional API outages that Github has). Curious to see your take on where Pierre is better.
Understood. I am looking for a side-by-side comparison focused on your feature set, not Github's. You answered it partially by calling out your focus areas. Github API reliability has been iffy for us, so it would be good to quantify the difference we can expect to get with you.
Sure – our API is built specifically for common LLM workflows. Here's a great example.
LLMs are often used for changing code. If an LLM creates a patch that touches 10 files, you need to take the following steps to save that patchfile on GitHub using their rest API.
.
1. Get the base branch SHA
2. Create a new branch (ref)
3. Create blobs (one per file… 10 blobs!)
4. Create a tree containing those 10 file changes
5. Create a commit
6. Update the branch ref to point to the new commit
7. Pull the GitHub api until it stops returning 422's (an eventual consistency issue when GitHub is under high load)
.
About 15 total requests…
With code.storage you just post the complete diff:
On top of ergonomics, we have first class APIs for git notes, grep, get archive (include/exclude by blob), and other filesystem behavior that is exceeding helpful when working with LLMs.
This is very nicely done. We have seen the same issue at a higher level of getting separators right when generating multiple files in a single inference call.
curious: wdym by "getting separators right when generating multiple files in a single inference call"
context: created hypertokens an even more robust hashing mechanism to create context-addressable memory (CAM), one cheat code is make them prefix-free, lots of others that get deep into why models work the way they do, etc.
These are large coal seams with significant exposure to the atmosphere. See https://en.wikipedia.org/wiki/Jharia_coalfield for an example. That excavator in the picture is not trying to put out the fire, it is just mining coal that happens to be burning. Spray some water, put out the fire and ship it off to customers.
The system prompt in this experiment limits the solution to always spell out the concrete moves verbally. A human solving the Tower of Hanoi gives up around N=4 and goes off to invent a recursive solution instead. Prompted differently, the LLM would solve these puzzles just fine.
Gary Taubes has written multiple scholarly books on the subject. He has won the Science in Society Journalism Award of the National Association of Science Writers three times. I think a training in one scientific field does qualify a person to go spend their time digging into other fields. How well they do is a matter of personal ability and effort and one would have to read their output and judge for oneself. I have read many of Taubes's book and many of the papers cited in his books and have concluded that he is a well qualified expert in the field.
