Hanchen Li
@lihanc02
Followers
152
Following
227
Media
9
Statuses
74
Year 0 PhD student at @UCB Sky Currently at @lmcache :) I work on AI Systems
Foster City, CA
Joined March 2022
Gave a pitch @genaisummitsf for the new one-click LLM inference that we are gonna release @lmcache next week. Thanks a lot for having us and stay tuned 🤫
1
1
2
Trying to get a learner’s permit this week. Cramming like a high schooler XD.
0
0
2
Pretty sure Grok will top out on “avatar arena”😏.
🚨 Breaking News: Grok 4's result is now live!. With 4k+ community votes, xAI’s Grok-4 tied for #3 overall in Text Arena — a huge leap from Grok-3. It scores Top-3 across all categories (#1 in Math, #2 in Coding, #3 in Hard Prompts). Detailed analysis in the thread 🧵
1
0
5
Our understanding of human intelligence will not surpass AI capabilities. For everything we can understand, we can build the digital simulation for it.
0
0
1
Getting an AI model subscription for the first time since 2025 :). (I canceled my OpenAI subscription after starting to use cursor)
1
0
1
It would be interesting to see different companies start to diverge and start optimizing for different workloads. Ex. chat models providing emotional support need not be the smartest but needs to echo with the user. Agentic models are better off being more concise.
1
0
0
Join our slack channel to get latest updates and share your feature requests with us! . Our developers loop through the channel every day to solve whichever problems you meet during the usage of this best kv cache management tool!.
🤔 Need help with LMCache? . Join our Slack! We are constantly checking for any questions people may have and would love to be of assistance!.
0
0
1
3) the ultimate goal for optimizing the flows should be how do we design a process such that humans can augment AI to finish the job quickly and accurately for the human requirements instead of building a machine to do everything (and complain when it does not).
0
0
0
2) people could argue that this is not true automation. However the truth is user won’t be able to do much else during this 60s period (it would even disrupt the flow if users shift focus).
1
0
0
Reasoning functionalities should always be optional in applications. 👉While the model could spend 60s thinking to give you a slightly better answer, it could also use the same time to give you an answer first, asks you for feedback, and iterate to have a better one. More below👇.
1
0
0
🏃♂️All the AI research labs used to be chasing the best "AGI model" in 2023. However, smaller companies are now specializing in specific domains to remain competitive. Recent examples include Devstral-medium-2507 (coding). Kimi K2 (agentic).
1
1
6
Submitted a paper to a non-system conference for the first time. 😲Got rejected with all positive reviews. Is it because AC opinion matters much more than the reviewers? .(Normally in systems there is a strong correlation between score and acceptance)
0
0
1
When 20-people startups got acquired for 700M no one is surprised. (sometimes just for the founders). However, when Meta spend 100M or 200M to just hire the guy for 4 years (and likely bringing some of his best man), it becomes a crime. Is there something that I am missing?.
Last week was 100m, and this week it's 200m. I'm waiting 8 more weeks for the 1B offer.
0
0
1