Eric Chen (@ErrkC) profile picture

Post-training lead @ Handshake Labs | Ex-Reka AI, Ex-Google Brain

United States
Joined July 2011
Photo by Eric Chen (@ErrkC)Photo by Eric Chen (@ErrkC)
Google should just sell TPUs directly.
66
Has Spirit Airlines thought about trading away their planes for GPUs and becoming an inference provider?
38
Use the new `auto-mode` they said, it will save you time they said ...
82
adhd is the real moat
I’ve had to learn to multitask 😅
1
2
1,085
Nice work @DaniYogatama and @RekaAILabs team!
Meet Reka Edge – Our next-generation vision language model for physical AI. Uses 3x fewer input tokens and achieves 65% faster throughput compared to leading 8B models. Image understanding, video analysis, object detection, and tool use. Built for Action. Fast enough for production, deployable anywhere. Read more: reka.ai/news/reka-edge-front…
3
122
Note all of these stats will still be true if even if Cursor had insane churn numbers: >  2x more people using agents in cursor than using the TAB feature. True if 90% users churned. > 90% of cursor users are daily active users of the agent product. True if 90% of users churned. > The agent product grew 15 x last year. True if 90% of users churned. 15x of 10 is 150. Without a base this is meaningless. > The cloud agent product is now responsible for 35% of merged prs in Cursor." Again, true if 90% of their users churned
Why it is total BS that Cursor is "dying"? " I just look at the numbers. 2x more people using agents in cursor than using the TAB feature. 90% of cursor users are daily active users of the agent product. The agent product grew 15 x last year. The cloud agent product is now responsible for 35% of merged prs in Cursor." @Mkclements What does no one see with Cursor that everyone should see @milesgrimshaw @aye_aye_kaplan @apartovi @jonas_nelle @jediahkatz
1
122
Maintaining the main road won't get you promoted. Ultimately it's a reward problem. The promo/reward structure needs to incentivize user experience over technical feats.
This is such a classic Google move.
1
9
2,406
Unpopular opinion: Researchers are disproportionately "feeling the AGI" with coding agents because it's much easier to vibe code research code.
35
You can do this in any city, just tell the investor "Let's grab a coffee then sign the term sheet"
SF is the only place in the world where you can walk into a coffee shop and walk out with a seed round filled.
38
I’ve been saying this for years. Most Japanese food reduces to five seasonings: shoyu, miso, mirin, sake, and dashi. Some of the most disappointing meals I’ve had were 3-Michelin-star (or 4.3+ tabelog rating) kaiseki in Kyoto/Tokyo.
I will not buy into the "Japanese food" propaganda Sushi and Ramen are the only passable Japanese dishes. Every other aspect of their cuisine is bland, gross or derivative
2
399
Using the exact same logic: Why would farmers replace workers with farm equipment like tractors? Pretty simple. Farming equipments don't need a wage, health care, time off, sick leave, social security, medicare, or unemployment ...
Why would corporations like Amazon replace workers with robots? Pretty simple. Robots don’t need a wage, health care, time off, sick leave, Social Security, Medicare or unemployment benefits. Maybe it’s time to tax robots & use the revenue generated to help working families.
2
91
From brief testing, cursor + GPT-5.2 extra high fast is severely underperforming Opus 4.5 in terms of agentic coding task completion, it also consumes more tokens.
1
297
The better business model is to sell insider information to frontier labs on what these breaking changes will be, so their competitors can waste money on compute training on deprecated code.
One of you AI companies better buy us next or I'm going to make a bunch of breaking changes and invalidate all of the code you generate.
1
290
Benchmarks not reflecting reality has been a problem for a long time. It got noticeably worse right after Gemini 2.5 Pro. In our internal evals (medium context, heavy instruction-following), GPT-4O is still dominating. The newer models since then have only gone downhill.
"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"
1
1
3
2,892
The only robotics company I trust at he moment. This is what a robotics demo should be.
Today, we present a step-change in robotic AI @sundayrobotics. Introducing ACT-1: A frontier robot foundation model trained on zero robot data. - Ultra long-horizon tasks - Zero-shot generalization - Advanced dexterity 🧵->
6
9
2
178
41,245
You know we are in a bubble when VCs are attending NeurIPS
1
152
Laughable take for anyone who has worked on a functioning voice agent in production. 1) Latency/function calling are solved problems for voice agents. 2) OSS models are now so strong that it’s Anthropic/OpenAI who are getting squeezed.
Hypothetical: 1. Anthropic ships a sub 200ms latency voice model with native tool calling 2. the entire category of customer support / voice agents for "x" collapses (decagon, cresta, eleven, vapi, giga). 3. Zendesk / Intercom/ Salesforce do basic integrations. Ledgers like Oracle, Salesforce, ServiceNow etc will probably be fine and integrate or acquire the top. LPs are unsophisticated - institutional money is desperate for AI exposure. Founders have been cashing out on secondaries for Series A/Bs bc they're aware
1
3
154
For the simple fact that they are furthest away from end users.
1
46
SF networking is overrated. Stop bragging about who you met and just build.
3
98