> To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy.
IIUC, this exploit was found without internet access, since that was the initial breakout. So either it was in the weights or found through probing?
> Q: Are the releases aligned with pre-training efforts?
> A: There used to be a time not that long ago, maybe half a year, distant past, where the models would align with RL runs or pretraining runs ... now the naming is by capability. GPT5 is a capable model; 5.1 is a more capable model
> I also think it’s important to notice that a lot of these challenges they happen with humans too. The concept of prompt injection isn’t that different from social engineering, right? When somebody calls in and says, “Oh, I forgot my password, can you just help me this one time?”
I wonder if the error propagation problem could be solved with a “branching” generator? Basically at every token you fork off N new streams, with some tree pruning policy to avoid exponential blowup. With a bit of bookkeeping you could make an attention mask to support the parallel streams in the same context sharing prefixes. Perhaps that would allow more of an e2e error minimization than the greedy generation algorithm in use today?
Having the data structures is nice and all, but using them is kind of painful. They are certainly second class.
Having to use accessor functions or destructuring macros instead of just a period or -> is often annoying too. The lack of syntax has cons as well as pros.
Sparse attention essentially combines 3 types of attention optimizations:
1. Compression of the query input vectors to reduce the size of the KV cache
2. Selectively computing uncompressed attention on a subset of tokens based on the compressed blocks with the highest attention scores
3. Using sliding window for local attention at full resolution
> Both Full Attention and sparse attention models are pretrained on 270B tokens of 8k-length texts, followed by continued training and supervised fine-tuning on 32k-length texts with YaRN to achieve long-context adaptation. Both models are trained to full convergence to ensure fair comparison.
> our experiments adopt a backbone combining Grouped-Query Attention (GQA) and Mixture-of-Experts (MoE), featuring 27B total parameters with 3B active parameters
Evaluated on MMLU, MMLU-PRO, CMMLU, BBH, GSM8K, MATH, DROP, MBPP, and HumanEval. NSA outperforms full attention on 7/9.
Beats out H2O, InfLLM, Quest, Exact-Top, and full attention on LongBench
Perfect retrieval on 64k needle-in-a-haystack
The CoT eval is less convincing, but outperforms the FA on AIME24.
Training speed of 2-9x vs. FlashAttention
Decoding speedup of 4-12x vs. full attention ["expected"? Didn't see comparison to other attention mechanisms]
Great to see this is alive and progressing! I believe Ohm started life in Alan Kay’s research group, to build a graphical OS and office suite in 10k lines of code. I found this talk immensely inspiring https://m.youtube.com/watch?v=ubaX1Smg6pY
My takeaway was that autocomplete, boiler plate, and one-off scripts are the main use cases. To use an analogy, I think the code assistants are more like an upgrade from handsaw to power tools and less like hiring a carpenter. (Which is not what the hype engine will claim).
For me, only the one-off script (write-only code) use-case is useful. I've had the best results on this with Claude.
Emacs abbrevs/snippets (+ choice of language) virtually eliminate the boiler plate problem, so I don't have a use for assistants there.
For autocomplete, I find that LSP completion engines provide 95% of the value for 1% of the latency. Physically typing the code is a small % of my time/energy, so the value is more about getting the right names, argument order, and other fiddly details I may not remember exactly. But I find, that LSP-powered autocomplete and tooltips largely solve those challenges.
The “pair programming” approach with good models is just slow enough that I lose focus on each step. The faster models I’ve tried are not good enough except for straightforward things where it’s faster to just use emacs/LSP refactoring and editing tools. Maybe supermaven manages to beat the “good enough, fast enough” bar; I’ll have to try it!
I still don’t understand how people are getting value out of AI coders. I’ve tried really hard and the commits produced are just a step up from garbage. Writing code from scratch is generally decent. But after a few rounds of edits the assistant just starts piling in conditionals into existing functions until it’s a rats nest 4 layers deep and 100+ lines long. The other day it got into a loop trying to resolve a type error, where it would make a change, then revert it, then make it again
ETA: Sorry forgot about the relevancy in my rant! The one area where I’ve found the AIs helpful is enumerating and then creating test cases
IIUC, this exploit was found without internet access, since that was the initial breakout. So either it was in the weights or found through probing?