Sam Cui

394 posts

Sam Cui banner
Sam Cui

Sam Cui

@35Kapc

ai/tech | the marketing guy | sharing actionable insights & opportunities

San Francisco Katılım Nisan 2014
146 Takip Edilen93 Takipçiler
Sabitlenmiş Tweet
Sam Cui
Sam Cui@35Kapc·
Trellis just crossed 10,000 stars! For most of last year, our team ran Claude Code hard and tried just about every workflow the community shipped — OpenSpec, plan-with-files, Superpowers — and none of them held up in real development. They're PRD-driven at heart: every new task, you rewrite the same architecture rules, naming conventions, and error handling from scratch. They struggle with anything project-specific, and even the custom rules you do define often never fire, lost to hallucination. So we built Trellis. The name comes from a garden trellis: we wanted to give code a structure — something that lets it grow fast while keeping its direction under control. We believe the better paradigm is Spec-Driven: using hooks to lock your hard-won engineering standards and workflows into every session, so parallel development and team collaboration finally feel smooth. From the first release, to the architecture overhaul in 0.5, to 20 back-to-back betas in 0.6 — across more than a thousand commits, we've reached this milestone hand in hand with you. Thank you to everyone who sat down with us for interviews, sent feedback, opened PRs, and helped answer other people's questions in the community. Thank you for backing us this early, while the product is still far from finished. Next, we're going after a new paradigm for how AI-native teams build software together — and on everything we've learned from Trellis, we intend to craft a product that truly earns the trust you've placed in us. Our story is just getting started.
Sam Cui tweet media
English
1
0
4
172
Floro S.
Floro S.@sflorimm·
Haven’t met anyone working less after adopting the AI.
English
101
11
205
15.3K
Sam Cui
Sam Cui@35Kapc·
@mobailabs the real tax is when both agents think they're right and you're stuck reconciling two valid but incompatible futures
English
0
0
0
1
MobaiLabs
MobaiLabs@mobailabs·
@35Kapc git worktrees per agent solves the file overlap. but the sneakier tax is divergent decisions — two agents refactor the same module differently and now you're merging two futures into one
English
1
0
1
6
Sam Cui
Sam Cui@35Kapc·
Running five Claude Code sessions at once feels like five times the output. The tax nobody mentions: every time you switch back, you reload the context in your own head, not the model's. That's where the speedup quietly leaks.
English
1
0
1
35
Sam Cui
Sam Cui@35Kapc·
@0xSero curious how much headroom is left in compression with more compute. feels like we're hitting diminishing returns but hope you prove otherwise
English
0
0
0
10
0xSero
0xSero@0xSero·
I'm spending 7k USD to make the best GLM-5.2 compression there ever was. Let's see how much better extra compute can make it. I am a huge believer in REAP / REAM / Ternary models and quantisation. If we're going to have an equal playing field compression has to get better.
0xSero tweet media
English
56
18
639
25.6K
Sam Cui
Sam Cui@35Kapc·
🛠️ Agent Reach — give your agent internet access without the API-key tax Your agent can write code, edit docs, run your project. Then you ask it to look something up online and it just whiffs. Reading a tweet, a Reddit thread, a YouTube transcript used to mean picking your poison: pay for each platform's API, fight the rate limits and logins, and babysit the scrapers every time they break — or paste it all in by hand. Agent Reach turns that into one command. One CLI, no API keys, no billing. It ships with whatever access method actually works for each platform right now and swaps it when things change — so you're not the one patching scrapers at midnight. What your agent can suddenly do: > pull YouTube + Bilibili transcripts and summarize them > search Twitter, Reddit, and GitHub — no keys, no bill > open a page and get clean text back, not raw HTML Trending on GitHub at 34.9k stars, MIT. An agent that can't see the internet is half an agent. github.com/Panniantong/Ag…
Sam Cui tweet media
English
0
0
0
19
Sam Cui
Sam Cui@35Kapc·
things that quietly kill your "five agents at once" speedup - forgetting what agent #3 was even doing - re-reading scrollback to find your own decision - two sessions that both edited the same file - the one you left "almost done" two hours ago - switching faster than any of them actually finish what's your real ceiling?
English
0
0
0
3
Sam Cui
Sam Cui@35Kapc·
@theohandsh all of the above, but distribution is the one that's hardest to fake
English
1
0
1
32
theo
theo@theohandsh·
AI is rapidly reducing the barrier to building software So what will separate the winners from everyone else? Product? Distribution? Brand? Community? Data?
English
18
0
15
1K
Sam Cui
Sam Cui@35Kapc·
@Dwriteway it compounds way faster than any title ever will
English
0
0
0
21
Kay
Kay@Dwriteway·
One digital skill can take you further than any job title ever could.
English
20
3
31
677
Sam Cui
Sam Cui@35Kapc·
Enjoy $200 coding plans while they still exist. These subsidies from model providers likely won't last forever. Thoughts?
English
0
0
0
16
Daniel Smidstrup
Daniel Smidstrup@DanielSmidstrup·
what actually makes a startup impossible to copy?
English
79
1
54
5K
Sam Cui
Sam Cui@35Kapc·
Hey founders, Looking to connect with people building in: - Saas - Tech - Automation - Ai tools - Web apps - Developer tools What are you building?
English
34
0
17
721
Sam Cui
Sam Cui@35Kapc·
@TTrimoreau distribution has always been more important, building is just cheaper now
English
0
0
0
3
Thomas Trimoreau
Thomas Trimoreau@TTrimoreau·
If everyone can build a startup in weekends now.. does distribution become more important than product?
English
55
0
50
2.3K
AshadBuilds
AshadBuilds@AshadBuilds·
@35Kapc I will be your first user I have seen quite a lot of websites/apps here, and i'm curious to try them. Feel free to share what you have built. I will drop a honest review about the product and how it felt using.
English
1
0
1
13
Shann³
Shann³@shannholmberg·
thought this was a solid idea, so why not take you down the rabbit hole of my thinking 1. LLMs are trained on what's widely available, so anything locked behind a course, a paid template or a private community, the model never saw it. that gap is what I want an agent to get into 2. but paywalled doesn't mean good. a lot of what gets sold is public knowledge repackaged, the same stuff everyone has with a nicer cover and a price on it 3. so this gets tricky for an agent. it can buy the course, but it doesn't have the taste to tell if what's behind the paywall is signal or repackaged, and judging that is difficult 4. still, it's a good experiment. take a skill like /last30days, the one that runs a whole sequence on a single command, and add CLI tooling for the places paid information lives 5. use a skill that creates a CLI for any site. point it at gumroad and it builds a CLI where an agent can search and buy anything there, then it does the same across multiple sources, skool, other directories, wherever the paywalled data sits 6. now it buys from all of them, pulls in information that was never in the training data, and turns it into a /llm-wiki, an obsidian brain on the topic I'm still working out the judgment problem, can an agent tell good from repackaged before it spends across all those sources, or do we need a human gate there
Shann³ tweet media
Shann³@shannholmberg

nvidia, stripe and nous teamed up on a hackathon for agents that can earn, spend, and run operations here are 4 I'd like people to build for marketing: > a media-buying agent for unconventional placements, the experimental stuff outside the mainstream platforms. newsletter sponsorships, niche site banners, podcast spots. it finds the placement, settles the price, and pays, and it only spends what you allow it to > an ops agent that sets up the whole environment for a campaign. it researches the toolset, decides on the stack, klaviyo for email, posthog for the analytics, then subscribes, pulls every api key and grants the access, so its ready to run with zero admin from you > a product-testing agent that buys and trials competitor products to map their flow. it signs up, pays, walks the onboarding, the emails, the paywall and the upsell, then writes up exactly how they move someone from signup to paid > a knowledge agent that buys its way up the learning curve. pick a topic and it pays for the courses and templates on gumroad, skool and anywhere else, ingests it, and turns it into an llm wiki your other agents can pull from

English
14
10
95
7.1K
Gergely Orosz
Gergely Orosz@GergelyOrosz·
Well this is ironic: Ironic: been recommending a resume service built by an ex-Dropbox eng for years (a side project, but a good one.) Dev pivoted to building an AI Engineer - fine! But now resume site is down. Customers billed. Support nonexistent. AI made it... a lot worse!
English
30
4
479
56.1K