CEO at https://teapawt.net/t.co/aKdnt7ep71
Cambridge, MA
Joined March 2010
- Tweets142
- Following295
- Followers97
- Likes123
Elad Ferber retweeted
Most people have no idea what it actually takes to be a founder. They talk about vision, grit, or passion. Those words are props.
What you really sign up for is a life where every decision feels like it costs something real. You will spend years being misunderstood. By your team, your family, even the people you hire to help you. You will fail in public and still need to keep the energy up in private. Every founder lives with the weight of knowing that you can do everything right and still get crushed by luck, timing, or somebody else’s mistake.
Founders aren’t braver than anyone else. They just get used to uncertainty, then stop waiting for clarity. Most of your wins won’t feel like wins at all. The first revenue will be too small. The first team will outgrow you or leave. The first product that feels right will barely matter to the market. You will doubt yourself in private, sometimes every week. The founders who last figure out how to keep moving while the ground shifts underneath them.
Most outsiders want the founder badge but none of the scars. They want the upside, not the drag. The hardest part is sticking around after every plan gets blown up and you have to rebuild with less optimism and more scar tissue. What makes it work isn’t relentless hustle or some mythical trait. It’s learning to make peace with constant discomfort, and then making decisions anyway.
If you need constant reassurance, you’ll give up before the real work begins. If you want everyone to like you, you’ll never make the calls that matter. If you can’t handle months where nothing feels certain, this life will eat you alive.
But if you can hold your own in chaos, get better at being wrong, and still want to show up and try again, you just might have a shot at building something that matters.
That’s what it actually takes. And nobody cares until you make it work.
Elad Ferber retweeted
🚀 AI is rewriting the rules of startup success. In our latest Crunchbase piece, Noah Lin and I dive into how we @defyvc have seen AI-first companies scale smarter with value model tactics—leveraging predictive power and network effects to reshape industries. 🚀
From defying traditional metrics to building data moats, these strategies are a game-changer for founders.
💡 Stay tuned for a deeper follow-up where we’ll share actionable tactics for founders to scale their AI ventures! Full article 👇
Elad Ferber retweeted
Context engineering is increasingly the most critical component for building effective AI Agents in the enterprise right now. This will ultimately be the long pole in the tent for AI Agents adoption in most organizations.
We need AI Agents that can deeply understand the context of the business process that they’re tied to. This means accessing the most important data for that workflow, using the appropriate tools at the right moment, having proper objectives and instructions, and understanding the domain that they’re in.
Some of the big open items for anyone building enterprise agents are:
* Narrow vs. General agents. The smaller the task, the easier it is to give the AI Agents the right context to be successful. But the smaller the task, the less value there will be. Finding the optimal task size for value generation will be an important factor for the next few years.
* Getting data into an agent-ready system. Enterprise data is often fragmented between dozens or hundreds of systems, many of which are not prepared for a world of AI. Most companies will still need to modernize their data environments to get the full benefit of AI Agents.
* Accessing the *right* data for the task is paramount. Even when you have data in a modern environment, getting access controls perfectly aligned to what the AI Agent is going to need access to is critical. Further, deciding what to do RAG on vs. just a general search vs. what to put fully into the context window will matter a ton per task.
* Choosing what should be deterministic vs. non-deterministic. If you demand too much from the models, you’re likely to see some drop off in quality. Yet, if you have the model do too little, then you’re dramatically underutilizing what’s possible with AI. This of course is a moving target because the models themselves are improving at an accelerating rate.
* The right user interface to get the AI Agents context deeply matters. Half of the problem for getting context to agents doesn’t look like an AI problem at all. It’s all about where the agents show up in the workflow and how the user interacts with them to provide them the context necessary to do the task.
The race for the next few years in AI in the enterprise is to see who best to deliver the right context for any given workflow. This will determine the winners and losers in the AI race.
Elad Ferber retweeted
Alexandr Wang: The future isn't humans vs AI - it's humans managing AI chaos. Even self-driving cars need 1 human operator per 5 cars to handle edge cases. The last 10% of accuracy? That's where humans shine.
Management isn't going away, it's evolving.
Elad Ferber retweeted
The best founders operate on a different clock - and it drives most of their employees crazy..
Elad Ferber retweeted
It was an absolute pleasure speaking with Winnie da Silva for the latest episode of her podcast, Transformative Leadership Conversations.
The title, "Dream Big, Iterate Fast" beautifully reflects what we discussed and the essence of how I view successful company building. Founders have to dream the dream while simultaneously learning fast and iterating accordingly to get there.
A few others topics that came up:
💡 The Pitch Shift That Works – Why some founders may undersell their vision without realizing it and how simple shifts in storytelling can make all the difference
🏗️ How to Pick the Right Startup to Join – Why choosing the right founder and team matters more than your initial role
🚀 Investing in Female Founders – Why women make exceptional CEOs and how they can sometimes undersell their vision
🏆 Lessons from Rowing & Leadership – What I've learned from rowing about choosing the right team, adapting in fast-moving startups, pushing past perceived limits and other leadership lessons I've carried from the water to VC
👀 Why I Hate the Spotlight – Just like my discomfort posting about this conversation, I'd prefer talking about my founders and being behind the scenes
🎧 Tune in via the link below or on your podcast platform of choice for our full conversation.
cc: @defyvc
Elad Ferber retweeted
# automating software engineering
In my mind, automating software engineering will look similar to automating driving. E.g. in self-driving the progression of increasing autonomy and higher abstraction looks something like:
1. first the human performs all driving actions manually
2. then the AI helps keep the lane
3. then it slows for the car ahead
4. then it also does lane changes and takes forks
5. then it also stops at signs/lights and takes turns
6. eventually you take a feature complete solution and grind on the quality until you achieve full self-driving.
There is a progression of the AI doing more and the human doing less, but still providing oversight. In Software engineering, the progression is shaping up similar:
1. first the human writes the code manually
2. then GitHub Copilot autocompletes a few lines
3. then ChatGPT writes chunks of code
4. then you move to larger and larger code diffs (e.g. Cursor copilot++ style, nice demo here youtube.com/watch?v=Smklr44N…)
5....
Devin is an impressive demo of what perhaps follows next: coordinating a number of tools that a developer needs to string together to write code: a Terminal, a Browser, a Code editor, etc., and human oversight that moves to increasingly higher level of abstraction.
There is a lot of work not just on the AI part but also the UI/UX part. How does a human provide oversight? What are they looking at? How do they nudge the AI down a different path? How do they debug what went wrong? It is very likely that we will have to change up the code editor, substantially.
In any case, software engineering is on track to change substantially. And it will look a lot more like supervising the automation, while pitching in high-level commands, ideas or progression strategies, in English.
Good luck to the team!
Today we're excited to introduce Devin, the first AI software engineer.
Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork.
Devin is an autonomous agent that solves engineering tasks through the use of its own shell, code editor, and web browser.
When evaluated on the SWE-Bench benchmark, which asks an AI to resolve GitHub issues found in real-world open-source projects, Devin correctly resolves 13.86% of the issues unassisted, far exceeding the previous state-of-the-art model performance of 1.96% unassisted and 4.80% assisted.
Check out what Devin can do in the thread below.
Elad Ferber retweeted
I think people have been whiplashed too much by founders who ping pong on strategy and decision making.
There's a new type of founder profile that resonates with operators:
- multiple startup experiences
- good at delegating
- only changes strategy w/ multiple quality signals
Elad Ferber retweeted
One magical part of Cursor’s internal tech stack is a prompt compilation library called priompt (github.com/anysphere/priompt)
Here's why works so well... (1/12)
Elad Ferber retweeted
1/3 - Our research, @npjDigitalMed today: #AI to predict risk of near-term #afib within 2-weeks from AFib-free single-lead ECG intervals (400,000 individuals). W MatteoGadaleta, @SteveSteinhubl , @leftbundle.
@scrippsresearch @ScrippsDTC @ScrippsRTI
rdcu.be/dtnP3
Elad Ferber retweeted
# On the "hallucination problem"
I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines.
We direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful.
It's only when the dreams go into deemed factually incorrect territory that we label it a "hallucination". It looks like a bug, but it's just the LLM doing what it always does.
At the other end of the extreme consider a search engine. It takes the prompt and just returns one of the most similar "training documents" it has in its database, verbatim. You could say that this search engine has a "creativity problem" - it will never respond with something new. An LLM is 100% dreaming and has the hallucination problem. A search engine is 0% dreaming and has the creativity problem.
All that said, I realize that what people *actually* mean is they don't want an LLM Assistant (a product like ChatGPT etc.) to hallucinate. An LLM Assistant is a lot more complex system than just the LLM itself, even if one is at the heart of it. There are many ways to mitigate hallcuinations in these systems - using Retrieval Augmented Generation (RAG) to more strongly anchor the dreams in real data through in-context learning is maybe the most common one. Disagreements between multiple samples, reflection, verification chains. Decoding uncertainty from activations. Tool use. All an active and very interesting areas of research.
TLDR I know I'm being super pedantic but the LLM has no "hallucination problem". Hallucination is not a bug, it is LLM's greatest feature. The LLM Assistant has a hallucination problem, and we should fix it.
</rant> Okay I feel much better now :)
Elad Ferber retweeted
The exciting thing about AI startups right now is that most are solving problems we could never do with software before. The battlefield is not merely startups vs. incumbents, but entirely new markets emerging that wouldn’t have existed without AI.
Elad Ferber retweeted
When you talk to more established founders, there's a common theme: late-stage investors pull some nasty shit. Once you get past series B you should stop optimizing for investors who can help you and switch to optimizing for ones who won't knife you if things go bad.
Elad Ferber retweeted
For a long while, you probably aren't a great CEO
But you know the customers
The software
The feature gaps
The roadmap
The competition
How to sell it
How to market it
How to build it
So maybe as flawed as you are
You are still the best CEO
Elad Ferber retweeted
Our portfolio company Synthpop and @eferber launched their first product today! Synthpop makes a set of connected AI assistants for healthcare workflow automation. They just raised $2.6M to help more heathcare systems.
prnewswire.com/news-releases…
Elad Ferber retweeted
A big part of being a CEO is sales. It starts with selling a cofounder on *you*, then investors on an *idea*, then customers on a *solution* to their problem, then employees on a *vision* they can be part of, and on and on it goes.
Selling is probably the most valuable yet underrated skill in tech.
Elad Ferber retweeted
Quick note – we've transitioned from the deprecated vicuna benchmark to a more advanced MT-bench, including more challenging tasks and addressing biases/limitations in gpt4 eval.
We find OpenChat's performance on MT-bench is similar to wizardlm-13b.
That's said, there remains a significant gap between open models and GPT-3.5, which is exactly what we aim to emphasize with MT-bench - to highlight this discrepancy. Though not flawless, it's one step towards a better chatbot evaluation.
Please check out our paper/blog for more technical details and leaderboard for complete rankings.
Blog: lmsys.org/blog/2023-06-22-le…
Leaderboard: chat.lmsys.org/?leaderboard
MT-bench Paper: arxiv.org/abs/2306.05685
I've been waiting for this 🤯
Open Source LLM Models Surpass GPT-3.5 🎉
In a groundbreaking development, a remarkable set of open-source LLM models has outperformed the capabilities of GPT-3.5.
What truly amazed me is not only the exceptional performance of these models but also the astonishing dataset they were trained on. Surprisingly, this dataset consisted of just 6,000 GPT4 conversations. 🤯
It's clear that high-quality data is working wonders. ✨
A few more details:
- OpenChat: Built upon LLAMA-13B, with a context length of 2048.
- Achieves an impressive 105.7% of the ChatGPT score on the Vicuna GPT-4 evaluation. 🏆
- Achieves a remarkable 80.9% win-rate on AlpacaEval. 🥇
- OpenChat-8192: Based on LLaMA-13B, with an extended context length of 8192.
- Achieves an outstanding 106.6% of the ChatGPT score on the Vicuna GPT-4 evaluation. 🌟
- Achieves an impressive 79.5% win-rate on AlpacaEval. 🥇
📚Links:
OpenChat on HF: huggingface.co/openchat/open…
OpenChat 8192 on HF: huggingface.co/openchat/open…
OpenCoderPlus: huggingface.co/openchat/open…
Dataset: huggingface.co/datasets/open…
Code: github.com/imoneoi/openchat
Elad Ferber retweeted
"Sam altman won't tell you that GPT-4 has 220B parameters and is 16-way mixture model with 8 sets of weights?"
@realGeorgeHotz is doing it again! 😂
