Sabitlenmiş Tweet
Are we actually facing diminishing returns from model scaling? What happens when we extrapolate scaling laws for the next few model build outs?
New essay (full link below):
English
Kevin Niechen
125 posts
Kevin Niechen
@kevinniechen
Principal @8VC. Previously @Neeva @Airbnb @Pinterest and systems research
San Francisco, CA Katılım Nisan 2016
483 Takip Edilen762 Takipçiler
It’s my birthday… and it’s my first week at @xai.
I have always believed that people are the real multiplier in any company.
Technology changes, markets shift, but the right team is what bends the curve.
That belief is why I joined @xai.
Founders who have shipped to millions.
Engineers who stay locked in (even for 36h) until the thing works.
Operators who treat impossible like a suggestion.
No theatrics. Just execution.
My role sits right in the middle of that energy as a talent engineer.
I am here to bring in world class people, support them, and build the systems to keep bringing people at full steam.
You need the strongest concentration of talent you can assemble to advance AI in a way that understands the universe… that place is at xAI.
👇If you are a founder or builder who has shipped something people truly love, I would love to meet you and see how we can partner - my DMs are open!
English
Kevin Niechen retweetledi
Congrats to @WisprFlow on their Series A!
Has been a privilege to be involved from the Seed
I initially invested because I heard that @tankots and @SahajGarg6 were the top students in their Stanford class. It tracked!
Incredible product, team, everything. Onwards!
@8vc
Wispr Flow@WisprFlow
Major update today: We just raised a $30M Series A from @MenloVentures 🚀 Our vision? Building Jarvis from Ironman – the voice interface for the future of AI How are we spending this money? On you, if you're a cracked engineer or a growth fiend Full investor list in 🧵
English
Kevin Niechen retweetledi
I’m thrilled to announce @OutsetAI's $17M Series A funding round, led by @8vc!
Since day one, @AaronLCannon and I were driven by a simple truth: Companies build better products when they actually talk to their users. YC drills "talk to your customers" into founders for a reason—it works. But scaling those conversations has always been challenging.
I’m proud that we've created a platform that enables companies to have deeper conversations with more customers at unprecedented scale.
What's even more special: not a single employee has left the company since we were founded three years ago. We believe the journey is just as important as the destination, and it's truly a privilege to work with our all-star team each and every day.
We approached Series A strategically, raising exactly what we needed to fuel our growth to the next stage. Thank you to 8VC and our other investors (Future Back Ventures by Bain & Company, @ycombinator (including our partner, the incredible @garrytan), @Adverbvc, @RebelFund_VC, Genius Ventures, Ritual Capital, Alt Capital, and Mute Ventures).
We're focused on what truly matters: serving our customers and growing responsibly. We believe in building a sustainable business that creates real value.
To our customers who trust us, our investors who believe in our vision, and our incredible team who make it all possible—thank you.
Time to scale.
English
Kevin Niechen retweetledi
We’re thrilled to share that @OutsetAI has raised $17M in Series A funding, led by @8vc to help companies better understand their customers.
When @mchlhess and I founded Outset, we set out to solve a critical problem: truly understanding your customer is painfully slow, expensive, and at times, impossible. It shouldn’t be.
We believed AI could change this by focusing on the human side of AI.
So we built the world’s first AI-interviewer. An AI agent that asked questions, listened, and learned from real humans.
By removing the barriers of traditional research and allowing companies to have deeper conversations with more customers more quickly, we could help teams infuse their products with genuine human understanding at a speed and scale previously impossible.
So what was the moment we knew we were onto something?
When WeightWatchers ran that first pilot back in early 2023. Participants shared incredible amounts of depth, all with the world’s first AI-interviewer.
We got a call the next day from their head of research: "You built a thing. We want that thing. How do I pay you for it?"
That validation has turned into explosive growth. We've doubled our revenue in just the last 4 months alone, with >50 enterprise customers like Nestlé, Microsoft, and WeightWatchers trusting us to unlock deeper customer insights.
All of this built on the backs of a truly special early team with 0 employee turnover.
This funding round was led by 8VC, with participation from an incredible group, including: Future Back Ventures by Bain & Company, @ycombinator / @garrytan, @Adverbvc /@aunder / @jess ), @_altcapital / @jaltma , Rebel Fund, Genius Ventures, Ritual Capital, and many more.
So what's next?
We're scaling our engineering and go-to-market teams to build the world's most powerful agentic customer research platform. All focused on one mission: helping companies better understand humans.
To our customers, investors, and team - thank you for believing in this vision.
Now check out the video to see Outset in action, testing the video itself with real people!
English
Kevin Niechen retweetledi
Privileged to lead @OutsetAI’s Series A and support them in making user research more scalable—and meaningful—than ever.
@AaronLCannon & @mchlhess have brought rare product chops, deep customer understanding, & impressive velocity to a $140 bn industry that has quickly emerged as one of the best use cases for an AI native approach we have seen to date
@kevinniechen and I wrote about how Outset is changing the equation (below):
Aaron Cannon@AaronLCannon
We’re thrilled to share that @OutsetAI has raised $17M in Series A funding, led by @8vc to help companies better understand their customers. When @mchlhess and I founded Outset, we set out to solve a critical problem: truly understanding your customer is painfully slow, expensive, and at times, impossible. It shouldn’t be. We believed AI could change this by focusing on the human side of AI. So we built the world’s first AI-interviewer. An AI agent that asked questions, listened, and learned from real humans. By removing the barriers of traditional research and allowing companies to have deeper conversations with more customers more quickly, we could help teams infuse their products with genuine human understanding at a speed and scale previously impossible. So what was the moment we knew we were onto something? When WeightWatchers ran that first pilot back in early 2023. Participants shared incredible amounts of depth, all with the world’s first AI-interviewer. We got a call the next day from their head of research: "You built a thing. We want that thing. How do I pay you for it?" That validation has turned into explosive growth. We've doubled our revenue in just the last 4 months alone, with >50 enterprise customers like Nestlé, Microsoft, and WeightWatchers trusting us to unlock deeper customer insights. All of this built on the backs of a truly special early team with 0 employee turnover. This funding round was led by 8VC, with participation from an incredible group, including: Future Back Ventures by Bain & Company, @ycombinator / @garrytan, @Adverbvc /@aunder / @jess ), @_altcapital / @jaltma , Rebel Fund, Genius Ventures, Ritual Capital, and many more. So what's next? We're scaling our engineering and go-to-market teams to build the world's most powerful agentic customer research platform. All focused on one mission: helping companies better understand humans. To our customers, investors, and team - thank you for believing in this vision. Now check out the video to see Outset in action, testing the video itself with real people!
English
Kevin Niechen retweetledi
Understanding customers is existential for any business. @OutsetAI is redefining how it’s done, with pioneering AI-led interviews blending survey scale & human depth. We’re proud to lead their $17mm Series A. More below from @jack_moshkovich & @kevinniechen:
English
Kevin Niechen retweetledi
After nine months in stealth, we’re excited to announce Augment has raised $25M, led by @8VC, and is launching Augie—the first AI teammate for logistics.
We are applying AI—the most transformative tech of our time—to logistics, the largest industry in the world.
🚀 We’re hiring! If you want to build AI that moves freight, let’s talk.
English
Kevin Niechen retweetledi
20+ years ago, @Meta & @PalantirTech formed in dorm rooms, enlisting bold young builders to create enduring companies.
AI revolutionized the toolset, but courageous top talent is still paramount. The @8VC Llama Stack Innovation Challenge is your invitation to be discovered.
8VC@8vc
Our friends @Meta are taking part in the 8VC Llama Stack Innovation Challenge! First challenge: 3/7-4/19. Use Llama Stack to build desktop, field, and edge AI applications that run locally for privacy and performance. Join now: llamastackchallenge.com
English
Our friends @Meta are taking part in the 8VC Llama Stack Innovation Challenge!
First challenge: 3/7-4/19. Use Llama Stack to build desktop, field, and edge AI applications that run locally for privacy and performance. Join now:
llamastackchallenge.com
English
We're incredibly excited to be launching the 8VC Llama Stack Challenge today for university students in the US!!
The first challenge is about building desktop, field, and edge AI applications using Meta's open source Llama models: 3/7 - 4/19!
Join now at llamastackchallenge.com
8VC@8vc
Our friends @Meta are taking part in the 8VC Llama Stack Innovation Challenge! First challenge: 3/7-4/19. Use Llama Stack to build desktop, field, and edge AI applications that run locally for privacy and performance. Join now: llamastackchallenge.com
English
Kevin Niechen retweetledi
Kevin Niechen retweetledi
Looking for a couple great founding engineers...
A diligent dwarf, Thorin, loves routing incoming comms to the right colleagues, and mining them to create replies, processes and alerts. He has a lot of cool tricks up his sleeve, and will save me hours per week; other friends agree they need his help ASAP.
We’re building something awesome around AI, email, and team processes at @8VC.
Our team co-founded the AI CRM used by 3500 firms (Affinity!), and after Palantir, we co-founded a lot of other mission-driven, billion-dollar companies — Addepar, Saronic, Resilience, OpenGov, Epirus, etc.
I know a billion-dollar product mission when I see it. Our internal technical and design teams are building, and we're looking for a couple more A+ engineers to work with us and build a generational company.
Thorin will be based out of our offices in SF and Austin, TX. You’ll have the opportunity to work directly with me and 8VC’s CTO.
If you are based and obsessed, or even just a really great engineer who wants a new challenge: come win with us. DM @8VC, or visit THORIN.COM.
English
Kevin Niechen retweetledi
Introducing Bespoke-Stratos-32B, our reasoning model distilled from DeepSeek-R1 using Berkeley NovaSky’s Sky-T1 recipe.
The model outperforms Sky-T1 and o1-preview in reasoning (Math and Code) benchmarks and almost reaches the performance of DeepSeek-R1-Distill-Qwen-32B while being trained on 47x fewer examples!
Crucially, we open-source the dataset (DeepSeek open-sourced the model, not the data). Let's work together on this exciting direction of reasoning distillation!
🧵More info and link to the blog below!
English
Kevin Niechen retweetledi
New blog post on the case for developing oral drugs for antibody targets.
In an era of antibody me-toos, this strategy stands to truly differentiate.
Full post below:
English
@Miles_Brundage that's my mistake, good catch
i wrote the sentence, then later added 100M gpus as a point of comparison, and forgot to update the spreadsheet and blog now. fixing now, thanks
English
@kevinniechen The "100M GPUs" thing is confusing. The sentence says it's publicly announced but there is no such announced datacenter/the spreadsheet just says it's an extrapolation
English
Are we actually facing diminishing returns from model scaling? What happens when we extrapolate scaling laws for the next few model build outs?
New essay (full link below):
English
@Calclavia is there a specific type of data youre thinking of like video, embodiment, code, etc? i think the q is the scaling laws for each of these domains and we have some early data for some
English
Text data from the web is just a tiny fraction - a summary - of the human experience.
We can argue web data is perhaps the high order bit of what there is to learn, but, in some sense, I think models barely scratched the surface on the space of things it could learn.
Kevin Niechen@kevinniechen
Are we actually facing diminishing returns from model scaling? What happens when we extrapolate scaling laws for the next few model build outs? New essay (full link below):
English
@zeeshanp_ @sea_snell yup i read charlies paper yesterday and thought it was great work; rly promising to see more research here
English
i think the discussion around predicting how models will improve with scaling pretraining is definitely improving. @sea_snell has a great paper on how to predict emergent capabilities by finetuning smaller models. another aspect of scaling that’s often overlooked is that fact validation perplexity may not be the best metric for measuring performance. perplexity can saturate but that doesn’t necessarily mean the model can’t have new emergent capabilities.
x.com/sea_snell/stat…
English
@DVarvanec17259 inference time scaling is great and prob a big part of the future given all the teams working on it now, but it's not the vision that was pitched for the past few yrs and it's unclear today what scaling will look like in that regime
English
@EthaiReubinoff @8vc i think that we're starting to see this with cos releasing more and more direct-to-user products outside of model capabilities
English
@kevinniechen @8vc It seems that if this is the case resources will be directed to higher efficiency inference and better user applications.
English