Johnna Liu

526 posts

Johnna Liu

Johnna Liu

@OvercookedJoJo

building agent safety infra @sponsiolabs , cs dropout @Cornell , exTMT IB @GoldmanSachs

San Francisco Katılım Ocak 2025
300 Takip Edilen2.3K Takipçiler
Sabitlenmiş Tweet
Johnna Liu
Johnna Liu@OvercookedJoJo·
In this revolution, we are still at 1995. We still need more valid infra for agent harness to onboard fully autonomous intelligence to run their internet and run all of our work. Of course this time everything moves much faster
English
1
0
2
1.2K
Shengkun
Shengkun@shengkun_ye·
six weeks ago, i quit my job. then things started moving crazy fast. we got into @fdotinc in our second week. started launching every day. got our first paying customer. our paying customers keep topping up. got our first check, then the second. every day there’s so much going on. all your energy and emotions get compressed. before you can react, the next thing hits. there are ups and downs. sometimes it’s exhausting. but i love it. i love that life can hold so much more in the same amount of time. and i love being a founder.
Shengkun tweet mediaShengkun tweet media
English
23
2
102
4.9K
Johnna Liu
Johnna Liu@OvercookedJoJo·
founders treat what they built like their own babies. that's also why, when we're picking which market to enter, we avoid the ones where the incumbents are still founder-led. very easy to out-execute professional management, especially those who've never taken extreme risks to build something from scratch as underdogs. but extremely hard to out-care an iron founder.
Garry Tan@garrytan

I went to BJ's Restaurant last night with my kids. The bathroom was disgusting. The front of house was kind but sloppy and slow. The food upset my stomach and I woke up at 4am this morning because of it. Whoever BJ is, they probably aren't a real person, because everyone acted like nobody's name is on the door. I studied the management history of BJ's. The original founders left after the seventh location. Then it was sold to their accountants. Then it went public. Then the CEO resigned last year after 19 years and was replaced by an interim board member from Darden Restaurants, who was then replaced by a "Chief Concept Officer" promoted to CEO. The CFO also quit. Roaches behind the takeout counter in Coral Springs. Rodent droppings and mold in the ice machine in Pembroke Pines. An "F" retention score on Comparably. Glassdoor reviews that say "management turnover is high... that should say quite a bit about the company culture." Seven layers of management between the person cooking your food and anyone who owns the outcome. General manager reports to area director reports to regional director reports to regional VP reports to SVP of Operations reports to the COO (who started in January) reports to the CEO (who started last year). 218 locations. Founders long gone. Managers rotate every 18 months. The kitchen is run by compliance checklists, not pride. A dirty bathroom is nobody's personal failure because it's nobody's personal restaurant. This is the stewardship crisis in America in one building. In Chinese restaurants, the 老板 (laoban) is there. He tastes the food. He watches the kitchen. His family's reputation is the business. The restaurant is clean not because of health inspectors but because his name is on it. Haidilao built a $30B hot pot chain with less than 10% employee turnover. Servers can give you free dishes without asking a manager. Why? Because they're treated like stewards, not interchangeable parts. The West replaced stewardship with professional management. MBAs who optimize spreadsheets for people they've never met. CEOs who've never touched the product they sell. Politicians who sign the bills and spend the people's money but never checked the money built anything that helped the people they claimed to care about. Founder mode isn't new. It's the oldest idea in Chinese business culture. We just forgot it. The best founders I fund at YC are natural stewards. They own the outcome. They're in the kitchen tasting the food. They care about the bathroom. Most of society's problems are a stewardship crisis. Not a lack of resources or technology or intelligence. A lack of people who give a shit because their name is on it.

English
0
0
3
604
Johnna Liu
Johnna Liu@OvercookedJoJo·
ai hallucinations set me off harder when i'm already pissed off
English
0
0
0
166
Johnna Liu
Johnna Liu@OvercookedJoJo·
@NotOnKetamine bro u can build a DD consulting firm and sell your services to VC and that will be huge
English
0
0
1
554
K.
K.@NotOnKetamine·
luel.ai raised $31.2M. The datasets, ARR, and jobs don't reconcile. Not a leak. Their own public product surface. Luel's datasets, ARR, and contributor jobs don't reconcile. 🧵 not-luel.vercel.app
K. tweet media
English
38
24
378
132.8K
Johnna Liu
Johnna Liu@OvercookedJoJo·
i desperately need more attention/followers on X to sell my products so - I am giving away my 15 YC Startup School referrals. What you'll get: > Jensen Huang & Sam Altman @sama fireside chat. > Access to 25k in credits (OpenAI, Anthropic, AWS, etc) > A network of top talent, ICML/Neurips/ICLR authors, Olympiad medalists. > After parties across SF hosted by YC alumni & partners > Be in San Franscisco for free* ($500 flight coverage) If you want it, let's connect 🤝 Tell me about what you are building & why you salute🫡 @PavanKumarNY
Johnna Liu tweet media
English
2
0
4
435
Johnna Liu
Johnna Liu@OvercookedJoJo·
both make mistakes, but it will be a non-sequitur to frame this as "both make mistakes, so both can be governed by self-written harness" failure modes are different. human errors sit inside a causal structure, with restrictions including embodiment, layered goals, and causal-not-statistical priors, while LLM errors are just sampling events on tokens, and patterns can be reweighted (that's why prompt injection can always be an issue). that's also why harness question isn't simply about who has the authorship of design the system, but also about a holistic mechanism that fits the failure mode. not intentional marketing but you can check this out: sponsio.dev . we built an external layer outside the LLM agents, based on mathematical provability and machine-checkable constrains. functional analog of the veto layer of human cognition, hard constrains over LLM's probabilistic token prediction.
English
0
0
0
47
Johnna Liu
Johnna Liu@OvercookedJoJo·
@kangminlee whats wrong with u fatherfucker? Any issues with genders here?
English
1
0
1
152
Johnna Liu
Johnna Liu@OvercookedJoJo·
@alexstauffer_ Hey Alex we are building provable behavioral safety for ai agents, backed by our Best Paper Award research. Early stage but soon can be a perfect match with Ramp’s AI agent products. Cover both pre-deployment formal verification and runtime enforcement. Wanna grab a coffee?
English
0
0
0
196
Alex Stauffer
Alex Stauffer@alexstauffer_·
Will be in SF until Wednesday. If you’re working on something cool, would love to meet!
English
31
1
131
19.6K
Johnna Liu
Johnna Liu@OvercookedJoJo·
@sama @pavankumarny I added ICML/ICLR because they are also good ML conferences (and because i got rejected by Neurips and only published ICML😇)
English
0
0
1
292
Daniel Green
Daniel Green@dgrreen·
The Sam Altman and @miramurati texts from the day he got fired from @OpenAI in 2023 just became evidence in the @elonmusk v. @sama trial. It felt like a meaningful moment in AI history, so I turned it into a musical. The lyrics are the texts.
English
107
199
1.8K
381.4K
Emmett Shear
Emmett Shear@eshear·
It’s an honor just to be nominated
English
219
252
9.8K
619.3K
JER
JER@lifeof_jer·
@OvercookedJoJo Very interested. Looks cool. Link to schedule demo?
English
1
0
0
30
Johnna Liu
Johnna Liu@OvercookedJoJo·
Stop worrying about coding agents deleting your codebase, prompt injection hijacking OpenClaw, or client privacy data leaking through agent tool calls. Introducing Sponsio - the open-source deterministic firewall for AI agents. - SOTA on agent safety benchmarks - <0.01ms latency (5,000–60,000× faster than LLM-as-judge) - ZERO LLM runtime cost Define policies in natural language, Sponsio compiles them into unbreakable, machine-checkable rules for agents. Cover both single tool call and multi-step temporal agent behaviors. Onboard with one prompt or two CLI lines. Go grab it.
English
5
7
28
1.9K
Johnna Liu
Johnna Liu@OvercookedJoJo·
@shengkun_ye sorry to hear that =( maybe worth trying it out with a simple prompt so your future codebase always survives
English
0
0
0
57
Johnna Liu
Johnna Liu@OvercookedJoJo·
@lifeof_jer guess you might be interested! Wish we'd shipped earlier so you wouldn't have to suffer. It's open source. try at sponsio.dev
English
1
0
0
147
Johnna Liu
Johnna Liu@OvercookedJoJo·
For coding agent users (Claude Code/Codex/Cursor): System prompts can't actually stop coding agents from deleting your codebase, because they are the same instance. Sponsio: formal methods, mathematically provable You write rules in natural language ↓ Sponsio compiles it into a small state machine ↓ each tool call updates the machine's state ↓ agent calls `volumeDelete` before any approval event ↓ machine flips to "violated" → effectively block before execution All machine-checkable, unbreakable. <0.01ms latency. No LLM in the path.
English
2
0
0
288
Johnna Liu
Johnna Liu@OvercookedJoJo·
For founders/engineers building agents for your clients: If your clients have any privacy requirements (e.g. no PHI to OpenAI), You'd otherwise spend hours writing redaction proxies and policy-as-code by hand . Onboard Sponsio via Claude Code/Codex/Cursor, your coding agents then effectively propose contracts and integrate in your agent harness. More unbreakable rules can also be added by natural language. <0.01ms latency because no LLM-as-judge.
English
1
0
0
139
Johnna Liu
Johnna Liu@OvercookedJoJo·
For OpenClaw users: Sponsio can effectively block prompt injection from malicious skills. Starter bundle for OpenClaw community is also available for day-one safety protection. You can also add custom rules in natural language to rule your lobsters as you wish. Sponsio will help compile them. All by formal methods. Machine-checkable and unbreakable. No LLM in the path.
English
1
0
0
155