Reflection AI Raises $2B to Challenge DeepSeek From the U.S.  ## Metadata - Author: [[Hard Fork AI]] - Full Title: Reflection AI Raises $2B to Challenge DeepSeek From the U.S. - Category: #podcasts - URL: https://share.snipd.com/episode/bdd55b42-24ec-49ed-a8a1-494bfbdcde53 ## Document Notes ## Highlights - **U.S. Open Source Torchbearer** - Reflection AI positions itself as an American open-source torchbearer after perceived abandonment by other U.S. players. - The company argues openness and collaboration drove past AI progress and should again shape frontier models. Transcript: Jaeden Schafer So they just tweeted and they said, you know, that they're, you know, building their next phase of reflection. They said, we're building frontier, open intelligence, accessible to all. Really, it feels not really like a dig at open AI, but basically it feels like reflection AI is taking up what open AI was supposed to do. It's kind of interesting because Elon Musk, you know, was quite vocal about roasting OpenAI, about not being an OpenAI company, but XAI also is not open source. And so a lot of people kind of thought maybe he would be building the open source AI company, which he's not doing. And it feels like the torch in America, at least, has been passed to Reflection because, you know, globally have Quen and DeepSeq in China that are doing that. So it feels like Reflection AI is the American response to this. So they've raised $2 billion to build, basically they're going to be creating LLM models. They said technology and scientific progress is driven by values of openness and collaboration. The internet, Linux, and the protocols and standards that underpin modern computing are all open. This isn't a coincidence. Open software is what gets forked, customized, and embedded into systems worldwide. It's what universities teach, what startups build on, and what enterprises deploy. So I think they're making a really compelling case here. ([Time 0:01:26](https://share.snipd.com/snip/0c4dea69-301b-4975-9ddd-8222cfb9d375)) - **Huge Raise Before A Product** - Reflection AI raised $2 billion at an $8 billion valuation despite not yet releasing a model or product. - The raise signals strong investor belief in team, mission, and future frontier LLM plans. Transcript: Jaeden Schafer So they've raised $2 billion to build, basically they're going to be creating LLM models. They said technology and scientific progress is driven by values of openness and collaboration. The internet, Linux, and the protocols and standards that underpin modern computing are all open. This isn't a coincidence. Open software is what gets forked, customized, and embedded into systems worldwide. It's what universities teach, what startups build on, and what enterprises deploy. So I think they're making a really compelling case here. We've seen that obviously OpenAI has gotten an insane amount of bad PR for being an open source company, getting a bunch of donations that were tax-free for their open source foundation, And then turning it into a for-profit company, taking $10 billion from Microsoft, all of the drama that came with OpenAI and then trying to switch into now a for-profit organization. And Sam Altman insisted it was the only way for him to be able to raise money and for him to be able to basically create AGI or a good outcome for the company. Now, it would appear that Reflection just raised $2 billion. This is an incredible amount of money. We've seen other big companies or other AI LLMs, usually with top tier talent, the co-founders of OpenAI have all gone on to raise at billion dollar valuations or billion dollar rounds Of funding. $2 billion is really impressive. And it's also considering there is no AI model. Like reflection has not put anything out yet. They're planning on their next model, but they don't have anything put out. So this $2 billion here is raised without basically the product that they're promising. ([Time 0:02:14](https://share.snipd.com/snip/227db3d4-da54-4527-ac39-afe4df2df240)) - **Use Open Models For Sovereignty** - Choose open-source LLMs when you need customization, sovereignty, or lower ongoing costs. - Use open models to avoid vendor lock-in and to build country- or enterprise-specific solutions. Transcript: Jaeden Schafer And I think building on top of it is a huge point, right? When you think about these open source models, who's the target audience? It's countries that want to build like their own AI models for their whole country or huge organizations that don't want to be paying open AI absorbent amounts of money. They want to be able to customize it the way that they want. So there's a lot of resources that can go into these. ([Time 0:04:39](https://share.snipd.com/snip/f3175b21-e615-4540-824e-d1f0d4a2a9bd)) - **Top Talent And Ambitious Compute Plans** - Reflection AI recruited researchers with DeepMind and other top lab experience and aims to train models on tens of trillions of tokens. - They claim capability to train mixture-of-experts models and plan a frontier release next year. Transcript: Jaeden Schafer So they said over the last year, we've been preparing for this mission. We've assembled a team who's pioneered breakthroughs, including Palm, Gemini, AlphaGo, AlphaCode, AlphaProof, and contributed to ChatGPT and Character AI, among others. They have a really strong team that is backed from Google DeepMind is where a lot of their researchers are coming from. TechCrunch did a whole article on this. And they talked basically, the thing that's interesting that I guess I haven't mentioned yet is that, so they raised this $2 billion. They did this at an $8 billion valuation and that is a 15X leap from their valuation just seven months ago. So they had a $545 million valuation seven months ago and now an $8 billion valuation. So originally when they launched, I know a lot of people are asking like, how did they even get started? How are they even raising this much money? They were just focusing on coding agents, autonomous coding agents. And I think they very quickly realized they have the like, in order to build an autonomous coding agent, they're basically have the talent to be able to build an LLM that could go and compete Head-to with OpenAI, theoretically. The most famous coding agent and product on the market is Claude Code right now. And so if they're competing with that, like Claude Code is just backed by Claude, which is a direct competitor to OpenAI and is competing at a very high level. So I think this is really interesting. This is going to be an open source alternative to OpenAI and Anthropic for sure, but also Chinese firms like DeepSeek that are doing this open source, they're going to be competing directly Open source. So this was actually started just last year in March by Misha Laskin, who was doing reward modeling for DeepSeek's Gemini project. And so this is a really, really powerful team. They only have about 60 people currently that are working for them. Most of those are researchers and engineers. You can imagine the team, after raising $2 billion, is probably going to scale out to a lot of different areas like go-to and other areas now that they have enough money. But what's interesting here is their CEO, who is Laskin, said that they have secured enough compute clusters and that they hope to release a frontier language model next year that's Trained on, quote, tens of trillions of tokens. ([Time 0:05:04](https://share.snipd.com/snip/8ab9ec1d-86ee-49be-9648-2dd19bf820b1)) - **From Coding Agents To General LLMs** - Reflection began with autonomous coding agents and pivoted when they realized the same tech could build general LLMs. - Misha Laskin led the effort after work on reward modeling and coding agents showed frontier capabilities. Transcript: Jaeden Schafer So originally when they launched, I know a lot of people are asking like, how did they even get started? How are they even raising this much money? They were just focusing on coding agents, autonomous coding agents. And I think they very quickly realized they have the like, in order to build an autonomous coding agent, they're basically have the talent to be able to build an LLM that could go and compete Head-to with OpenAI, theoretically. The most famous coding agent and product on the market is Claude Code right now. And so if they're competing with that, like Claude Code is just backed by Claude, which is a direct competitor to OpenAI and is competing at a very high level. So I think this is really interesting. This is going to be an open source alternative to OpenAI and Anthropic for sure, but also Chinese firms like DeepSeek that are doing this open source, they're going to be competing directly Open source. So this was actually started just last year in March by Misha Laskin, who was doing reward modeling for DeepSeek's Gemini project. And so this is a really, really powerful team. They only have about 60 people currently that are working for them. Most of those are researchers and engineers. ([Time 0:05:42](https://share.snipd.com/snip/0df6a268-de7a-44c2-9bb6-e9f0299658f9)) - **Mixture-of-Experts Is A Core Bet** - Reflection emphasizes mixture-of-experts (MoE) architecture as a performance multiplier for specialized reasoning. - MoE lets a model route queries to specialized sub-models, improving scores on complex benchmarks. Transcript: Jaeden Schafer With this milestone unlocked, we're now bringing these methods to the general agentic reasoning. Yeah. Basically, like if you can do coding, they chose one of the hardest problems. And once they realized they could crack that, they're like, well, actually the technology could just be used as a general purpose LLM. Coding is a really good one because it's like thinking, it's logic, it's putting a lot of complex math and things together. And all of a sudden, if you can really crack autonomous coding, you're like, okay, well, I guess it's also just an LLM. I think that's what they realized. MOE that they're doing is specifically kind of the big breakthrough that DeepSeek had, which was followed by Quen and Kimi. Those are other models from China. But it's this mixture of experts. So basically, you ask the model a question, and it has these experts inside of the model. I think OpenAI has like 16 and it basically picks which of these 16 experts, maybe a coding expert, maybe a PhD in psychology expert, right? Like you can imagine these different experts and it will, the model will pick which of these experts are the best at, you know, fine-tuned and answering those questions. And so it seems like they've done that. And this is kind of what really enables these models to think much better. ([Time 0:07:46](https://share.snipd.com/snip/0d4a7c0a-e1b6-4cf4-b084-9e07efc096c5)) - **Geopolitical Framing Of The Mission** - Reflection frames its mission as preventing global AI standards from being set outside the U.S. - The pitch stresses that allies need an American open-source alternative to Chinese and other models. Transcript: Jaeden Schafer So what he specifically said, though, is that DeepSeek and Quinn and all these models are a wake-up call because if we don't do anything about it, then effectively the global standard Of intelligence will be built by someone else. It won't be built by America. So a really big part of their pitch is that in the United States, we are falling behind on open source, basically because OpenAI abandoned it. Now, is it true? OpenAI did not completely abandon it. They did recently release a really powerful OpenAI model, but I think a lot of people are worried that these are far and few, you know, these are few and far in between. Like open AI releases one and then like two or three years later, they might release another one because everyone's kind of harassing them and they want to get some goodwill. That's the way it feels. I don't know if this is accurate, but I'm just telling you how it feels. And so if there's a company that's exclusively focusing on it and this is their mission, I think that a lot of people will be willing to support them and back this. And in addition to this, it puts the US and all of our allies at a disadvantage because right now enterprises in sovereign states, they're not going to use the Chinese models due to a lot Of different legal repercussions. And so if we're not using those, we need an alternative. And so this is kind of their pitch to become the alternative. And specifically, he said, quote, so you can either choose to live at a competitive disadvantage or rise to the occasion. So sets himself up as this. ([Time 0:08:56](https://share.snipd.com/snip/0ec8ca88-58d4-4210-9354-ca725e05d439)) - **Broad Industry Support** - Industry figures welcomed Reflection as a much-needed American open-source entrant to offer cost, customizability, and control. - Hugging Face and White House advisors framed the move as strategically important for U.S. competitiveness. Transcript: Jaeden Schafer Clem DeLang, who's the co-founder and CEO of Hugging Faces, which is kind of an open source collaboration platform for AI builders. He was talking to TechCrunch about all of this. And he said, quote, this is indeed great news for American open source AI. Now the change will be to show high velocity of sharing of open AI models and data sets, similar to what we're seeing from the labs dominating in open source AI. ([Time 0:11:04](https://share.snipd.com/snip/7d5ab6de-a35d-46e4-bb94-9fad17750ff6)) - **Investor Signal: GPUs And Scale** - Major investors including NVIDIA and top VCs backed the round, signaling hardware and capital alignment. - The investor list reinforces expectations Reflection will spend heavily on GPUs and scale quickly. Transcript: Jaeden Schafer So as far as who's actually investing and putting money in, they have NVIDIA, of course, why would NVIDIA not put money into a $2 billion, you know, an $8 billion company that's taken In $2 billion that inevitably is going to have to spend most of that on NVIDIA GPU chips. So great move by NVIDIA. It's also being invested in by Disrupt, DST, 1789, B Capital, Lightspeed, GIC, Eric Wang, Eric Schmidt, Citi, Sequoia, CRV, and others. So basically all of the big players that have been invested in all the big AI companies are jumping onto this really high signal and obviously has a very impressive team that is putting This together. Laskin is a legend. And so I think he's able to raise that kind of legendary money. ([Time 0:11:32](https://share.snipd.com/snip/69d1b6d9-f9f4-4ce7-bf58-4c5f2550dd6b))