weilei (@WeileiY) profile picture

Building something, led growth and GTM of two unicorns from seed, onto the second act.

New York, USA
Joined June 2012
Honestly the only reason Toly is spending so much energy debating against Arb/RH chain is because they are becoming a threat. Take it for what it might mean.
Lmao. The spreads are worse on arb rn and the fees are higher. Arbs 10% rake from the fees alone is more bips than the sandwich rate, and that鈥檚 not even accounting for the worse spreads. By a factor of 10x from what I can tell. A single sequencer maximizing shareholder value will never beat permissionless competition sandwiched.me
1
3
18
1,523
The biggest blunder Anthropic made was giving people a reason to try Codex and GPT 5.5. I used to be a die-hard CC user, but after trying out Codex for code implementation (using CC for planning, review, and E2E tests), it is unlikely I'd ever go back to a single provider.
I'm having a lot success using Fable xhigh as a planner/architect, using GPT 5.5 xhigh (subscription) as a coder, then Fable xhigh again as a judge. At API pricing, planning+judge costs are in the ~few dollar range compared to typical $50+ full round trips. I've seen some others using dumber/cheaper coders, but GPT 5.5 even at xhigh compared to Fable 5 is very cheap and very fast. And GPT 5.5 is just... really good. Still been less than 24hrs since the re-release so the longevity of this approach is unclear, but its been working really well.
3
1
7
593
A while back I ran a series of eval scenarios, Sonnet 4.6 was already heck of commercial model when you think from a successful task completion/efficiency point of view. $/M token is very limited way of measuring. DS 4 got better but it was before Sonnet 5.
Hot take: Unsure who鈥檚 the target audience for that? Chinese models likely better at fraction of price. We want frontier to deliver frontier not mid. Will test and (compare with glm) and report back if first impression was wrong
3
4
539
When you do DeekSeek V4 vs Sonnet 4.6 it was pretty close neck to neck, but now that Sonnet 5 is out I need to re-run the evals.
194
Cash for Cache - 70% cache hit rate for $100/user cost is a whopping $63/user saved per month! Context caching strategy for your agentic product can save or destroy your startup overnight. Anthropic, the leader in enterprise AI adoption, understood this very well and that is why they designed their APIs so cache read (of content you proactively cached during the first turn of a session) costs only 10% of non-cached input. That is the difference between profitability and bleeding for each added user. But that is only half of the story. How you build your cache blocks around carefully demarked cache points, such that you put heavy, cross-agent prompt/context blocks in the first blocks, and dynamically generated ones after so that you benefit from continuous cache saving across the largest set of agents and users. Fun fact, if you look under the hood, you can tell which use cases each LLM provider is prioritizing or specializes in based on how their caching logic and pricing works. Gemini is totally different from Anthropic, breakdown in a follow up post.
3
3
339
AI agents design pattern -> use deterministic approach whenever you can. If one agent can't interpret the signal correctly, you can bet two agents will have even harder time agreeing.
2
1
2
301
A hard lesson every agent builder face is assuming adding live voice conversation to your existing voice mode/STT is marginal effort. Wrong, be ready for exponential complexity. On voice mode all you have to do it handle speech-to-text and feed back to your LLM as normal prompt. You can improve accuracy, remove filler words, pre-stream partial audio to speed up conversion, etc. All quality of live improvements for sure. However, once you step into the Live Voice mode, it is a completely different beast. A fitting comparison is the difference between downloading a full video from torrent to watch later vs streaming it live. With live voice you have to deal with, but not limited to: - Voice Activity Detection (VAD)聽to detect in real time when a speech supposedly starts and ends, to avoid awkward prolonged silence. There are some frameworks but be ready for heavy tweaking. - Continuous token streaming with usually "isFinal" flag from STT providers, but be aware isFinal might not be truly final if user jumps in again to add more context - You want to trigger LLM inference asap, but also have hooks ready to shut down any in-progress inference if user jumps back in with more context, and loop continuous - Then add the text-to-speech provider so user can hear a human sounding voice, which is just your LLM output. There is the whole art of adding modifier tags to the sanitized output (even more fun with latest Gemini TTS). Ah be ready to also interrupt this stream if user jumps in at any time. - You also need to deal messy transcript over-growth issue if conversation lasts longer than a few minutes all the while trying to maintain a sub 500ms latency for a natural human conversation, which means the latest state of art LLMs are often too slow for but flash models are not nuanced enough Happy building everyone. Feel free to share your live conversation journey or nightmares :D
2
1
2
327
Can open-weight models perform as well as Sonnet 4.6? The answer is Yes, but at steep cost, literally. I ran a comprehensive eval suite comparing Sonnet 4.6 against GLM 5.1, DeepSeek 3.2, and Qwen 3.6 (this one is proprietary though). Only DS 3.2 had a fighting chance.
2
577
While DS 3.2 outperformed Sonnet 4.6, it cost 2x more in $$ terms with a naive/no-cache implementation. I was puzzled because DS runs at $0.56/$1.68 per M tokens vs Sonnet's $3/$15. DS used 2.4x more turns on average (6.0 vs 2.5). Sonnet is still one-shot king in accuracy.
1
349
Super annoying Claude Code behavior with testing is it often leans toward loosening assertions instead of fixing the underlying cause. You have to be very explicit.
226
My latest @nba_topshot video - how to buy and sell your moments on the new marketplace! #NFTs #NBA #NBATopShot #flowblockchain youtu.be/2I3VVLhrUe8
2
3
3
9