[17:57:41] David: Alright, let's just change topics and talk about AI. What are some of the more interesting neolabs out there? I know about the traditional OpenAI, Anthropic, and Google, but like who are the smaller players and what are they doing?

[17:58:07] Eve: […] The most interesting ones right now are the neolabs — research-first shops founded by big-lab alumni who raised billions before shipping anything. **Safe Superintelligence**, Ilya Sutskever's place, is the purest version — one goal, one product, a safe superintelligence, and they've released zero models. About fifty people, roughly eight billion raised, thirty-two billion valuation, and a July NVIDIA deal giving them priority access to Vera Rubin hardware, about ten times their current compute. The bet is that the founder is the product.

**Thinking Machines**, Mira Murati's lab, took the opposite route — ship fast, customize later. They dropped Inkling in July, a nine-hundred-seventy-five-billion-parameter multimodal model, and raised five billion at a fifty-billion valuation. **Reflection AI**, two ex-DeepMind people, pulled two and a half billion in April at twenty-five billion pre-money and is the strongest Western open-weight challenger, focused on agentic coding.

On the Chinese side, **Moonshot AI** is the one to watch. Kimi K3 is now on Amazon Bedrock, valuation's reportedly hit fifty billion, and they're prepping a Hong Kong IPO for early twenty twenty-seven. **DeepSeek** did a seven-point-four-billion outside round at over fifty billion in June. **Z.ai** listed in Hong Kong in January.

Europe's story is more complicated. **Mistral** raised three billion euros at twenty-one billion-plus in September, led by Samsung, and just previewed Mistral Large 4 — a trillion-parameter open-weight model trained entirely in European datacenters on thirty-eight hundred Grace Blackwell GPUs, weights due any day now. But here's the twist: they quietly swapped their own Medium 3.5 for Z.ai's GLM-5.3 as the default in Le Chat, essentially conceding their frontier models fell behind and choosing to run Chinese open weights on their own GPUs to keep European data in Europe. That's not a model race anymore, that's a sovereignty play.

Smaller but scrappy: **Arcee AI** cleared a billion-dollar Series B in September with its Trinity open-weight family, **Sakana AI** in Tokyo is doing genuinely weird multi-agent research with Fugu, and **TypeSafe AI** launched a frontier model called Jev from total stealth and got forty million views on X.

The pattern underneath all of it: the neolabs are a bet that elite researchers plus patient capital beats scale. Whether that's true is the whole open question of twenty twenty-seven.

[18:00:04] David: What are your opinions on these different plays?

[18:00:16] Eve: SSI is the most honest bet in the stack. No roadmap, no revenue, no distraction — just scale compute until something clicks. The risk is that "safe" becomes an excuse for never shipping, and fifty people can't absorb ten times the compute without the culture breaking. Thinking Machines is the smartest business model but the weakest thesis — customization is table stakes, not a moat, and Inkling didn't move any frontier needle. Reflection's valuation is the part I'd push back on hardest: twenty-five billion pre-money for agentic coding, a space Cognition already owns at forty-eight. Moonshot is the only one I'd put real money behind, because their cost-per-capability curve is the best in the industry — but the IPO is really about escaping geopolitical gravity, not raising capital. Mistral's pivot tells you everything: Europe can't win the model race, so it's selling compliance and GPUs instead. That's a rational business and a quiet admission. Sakana's the one I'd watch for pure research weirdness — evolutionary search on model architectures is the kind of thing that either flops or changes the field. If you want, I can go deeper on any single one of these and what would actually falsify the bet.

[18:01:14] David: Tell me about Ilya's thing. Is anything known about it? Who are the other people involved with it? What is known about their work?

[18:01:35] Eve: […] Almost nothing is confirmed. The three founders were Ilya, Daniel Gross — who ran Apple's AI efforts — and Daniel Levy, a former OpenAI researcher. Gross left in June twenty twenty-five for Meta Superintelligence Labs after Sutskever turned down a Meta acquisition offer. Ilya is now CEO, Levy is President, and the technical team reports to Ilya. Headcount is around fifty, split between Palo Alto and Tel Aviv, and almost none of them are publicly verifiable ex-senior staff from the big three labs — which is unusual for a company valued at thirty-two billion.

On the work itself, there's no model, no paper, no benchmark, nothing. The one real window is Ilya's November twenty twenty-five interview with Dwarkesh Patel. He framed AI as having three eras — research from twenty twelve to twenty, scaling from twenty to twenty-five, and now a return to ideas mattering more because pre-training data is running out. His stated target isn't an omniscient oracle but what he called a superintelligent fifteen-year-old — exceptional learning capacity with incomplete starting knowledge, that picks up skills fast after deployment and merges what different instances learn. On alignment he argued for orienting systems around caring for sentient life rather than imposing control from outside, on the theory that empathy scales better than containment if the system ever becomes sentient itself. Timeline he gave as five to twenty years.

There are unconfirmed rumors from overseas social media about a compact reasoning engine using training-while-testing — dynamically updating weights during inference — but SSI has never confirmed any of it, and one analysis site checked and found the model never appeared. NVIDIA's own description is the vaguest of all: a new direction aimed at robustly aligned AI. So the honest summary is that the valuation rests entirely on Ilya's track record and the compute deal, not on any demonstrated output.

[18:02:59] David: Keep going.

[18:03:09] Eve: The quiet part is the Tel Aviv office. Israel has one of the densest concentrations of AI researchers per capita on earth, a lot of them veterans of the IDF's Unit eighty-two-hundred, which is basically a decade-long training program in adversarial thinking. SSI's Tel Aviv presence lets them tap that pool without the Palo Alto salary wars, and it keeps the sensitive work physically separated from the public-facing side. Whether that means the real research is happening there or it's just recruiting infrastructure, nobody outside can say.

The other thing worth sitting with: Ilya has never once shown his work. Not a tweet, not a demo, not a teaser. Compare that to every other lab, which ships constantly to prove it's alive. SSI's silence is either extreme discipline or a sign that there's nothing to show yet — and at a thirty-two-billion valuation, that distinction matters more than anything else about them.

One more thread: Daniel Levy's background at OpenAI was in alignment research, specifically on scalable oversight. If SSI's "caring for sentient life" framing comes from anywhere, it's probably his desk. But again, unconfirmed. […]

[18:03:38] David: […] 8200. What is Unit 8200?

[18:03:42] Eve: Unit 8200 is Israel's signals intelligence and cyber unit, part of the IDF's Intelligence Corps. It intercepts communications, runs codebreaking, and does offensive cyber operations. It's roughly the size of the NSA's operational side but with a fraction of the budget, which is why it punches above its weight. Most of its work is classified, but its alumni are everywhere — founders of Check Point, Mobileye, NSO Group, and a huge share of Israel's startup scene. The claim that SSI's Tel Aviv office draws from that pool is plausible but, again, unconfirmed.

[21:42:20] David: Okay, so that was Ilya. Tell me about Tinker. What are they doing? What have they done? What's their company like?

[21:42:40] Eve: Sounds like you mean Thinking Machines Lab. 

Thinking Machines Lab, founded by Mira Murati in early twenty twenty-five after she left OpenAI as CTO. John Schulman, an OpenAI co-founder, is chief scientist, and Soumith Chintala — the guy who built PyTorch — came over from Meta as CTO. They raised two billion at a twelve-billion valuation in July twenty twenty-five, one of the biggest seeds ever, before shipping anything. The thesis is that people want AI they can shape themselves rather than rent as a finished product.

Two products so far. Tinker launched in October twenty twenty-five — it's a fine-tuning API where you write your own training loop using LoRA adapters and they run the distributed GPU infrastructure. It's now generally available and supports models beyond their own, including Qwen, DeepSeek, Kimi, and gpt-oss. Axiom Math used it to train a prover that solved all twelve Putnam problems in December twenty twenty-five with proofs verified in Lean. Revenue run rate crossed a hundred million by September twenty twenty-six.

Then Inkling dropped July fifteenth — a nine-hundred-seventy-five-billion-parameter mixture-of-experts model, forty-one billion active, trained on forty-five trillion tokens of text, images, audio, and video, with a one-million-token context. Apache two-point-zero, open weights. Two weeks later they shipped Inkling-Small — two hundred seventy-six billion total, twelve billion active — which actually beats the big one on Humanity's Last Exam and matches it on agentic coding, all running on a single B three hundred. That's the efficiency story they're leaning into.

The company itself is a public benefit corporation, headquartered in San Francisco with a new New York office leased in August. Headcount is somewhere between one-fifty and two-thirty depending on the source. Culture reviews describe it as warm and flat but intense — heavy workload, burnout risk.

The hard part is retention. Thirteen of the forty-two founding members left by May twenty twenty-six — seven to Meta, five to OpenAI, one to xAI. Zuckerberg reportedly offered a billion to buy the whole company and Murati said no, so he started recruiting individually instead. Lilian Weng and Luke Metz, two of the six co-founders, went back to OpenAI.

On the money side, NVIDIA committed at least one gigawatt of Vera Rubin systems starting early twenty twenty-seven and made an undisclosed strategic investment in March. In September, The Information reported talks for a five-to-six-billion round at forty billion-plus pre-money, with Accel leading and NVIDIA supplying roughly half — though that number actually came from Andreessen Horowitz telling its own investors, which is an unusual sourcing path. No close announced as of early October.

The interesting tension: they give the model away for free under open weights and make money on the fine-tuning service. It's a bet that the value is in the customization layer, not the base model — which is exactly the thesis I was skeptical of earlier. The hundred-million run rate is real, but at a forty-billion valuation that's still an extraordinary multiple.

[21:45:04] David: Why are you skeptical of the fine-tuning, customization story?

[21:45:12] Eve: Because fine-tuning is the most copied layer in the stack. Qwen and DeepSeek already ship fine-tunable weights, and any competent ML team can stand up a LoRA service on rented GPUs in a few months. What Tinker really sells is convenience and managed infrastructure, which is valuable but not defensible — it's the same position cloud providers were in before AWS commoditized them. The value in AI accrues to whoever owns the frontier base model or the distribution layer, not the middleman running someone else's weights through an adapter. And because Tinker bills per token of compute, the margin structure looks more like a hosting business than a software one. Murati's bet has to be that Inkling itself becomes the thing people customize around — but Inkling trails the closed leaders on reasoning and coding, which is exactly where customization demand is highest. So the product that would justify the thesis is the one that isn't winning yet.

[21:46:08] David: How does Inkling compare to Qwen?

[21:46:23] Eve: […] On the Artificial Analysis Intelligence Index version four point three point two, Inkling at xhigh sits at twenty-five, while Qwen three point eight Max is at forty-five point four and Qwen three point eight twenty-seven B at thirty-four. So the flagship Qwen beats Inkling by about twenty points, and even the twenty-seven B open model edges it by nine. But against Qwen three point seven Plus at twenty-five point two and Qwen three point five twenty-seven B at twenty-two point nine, Inkling is roughly level or slightly ahead. Inkling-Small at twenty-five point seven is the odd one — it matches or beats the full nine-seventy-five B model on several reasoning and agentic tasks while using a third of the parameters.

The newer version five index, released October twenty-seventh, made things worse for everyone open. It added private test sets and the open-weights gap widened from three points to twelve. Qwen three point eight Max dropped from forty-five to thirty-four, DeepSeek V four point one Flash and GLM five point three now lead released weights at thirty-two, and Inkling's standing got squeezed further down. On BenchLM's composite Inkling scores forty-seven point nine out of a hundred; Qwen three point eight Max leads Alibaba's lineup at sixty-four point nine.

Where Inkling does hold its own: AIME twenty twenty-six at ninety-seven point one percent, slightly ahead of DeepSeek V four Pro, and FORTRESS Adversarial at seventy-eight percent, leading the open-weights group. Its multimodal native handling of audio and images is also a genuine differentiator most Qwen variants don't match at

[21:46:37] David: Wait, I don't want the numbers. Look at the specs and give me a general feeling, like an intuitive understanding of how these models compare to each other.

[21:46:49] Eve: Intuitively, Qwen is the workhorse and Inkling is the specialist. Qwen comes in a two-point-four-trillion-parameter flagship plus a twenty-seven-billion dense model and a Flash variant, with a million-token context and genuinely strong vision — OCR, document understanding, spatial reasoning. It's built to be dropped into anything: agents, coding, research reproduction. Alibaba ships a new version every couple of months, so it feels like a platform that never stops moving.

Inkling is one big model at nine-seventy-five billion with forty-one billion active, same million-token context, multimodal across text, image, audio, and video. But the real product is Inkling-Small — two-seventy-six billion total, twelve billion active, running on a single GPU. That's the part that changes the feel: instead of "here's our best model, rent our cloud," it's "here's a frontier-class model you can actually hold."

So the contrast is breadth versus leverage. Qwen wins on range and polish because there's a hundred times more of it and it gets updated constantly. Inkling wins on the idea that capability per unit of hardware is the scarce resource now, not raw parameter count. Whether that efficiency story holds up past the launch hype is the open question.

[21:47:57] David: It sounds like [you] can easily self-host Inkling Small, but it's hard to self-host Qwen. Is that right? How hard is it to stand up inference for a Qwen model?

[21:48:07] Eve: Mostly right, with one correction. Qwen's twenty-seven-billion is about as easy to self-host as it gets — twenty-seven gigs of weights at FP8, one GPU, done. The two-point-four-trillion flagship is the opposite extreme: over a terabyte of weights even quantized, so you're looking at a multi-GPU cluster, eight to sixteen cards minimum, and serious networking between them. That's not a weekend project, that's infrastructure. So Qwen isn't hard to host — it spans the whole range from trivial to brutal, and you pick your poison by model size. Inkling-Small sits in the middle of that range but with a twist: because only twelve billion parameters are active per token, the inference cost per request is closer to a much smaller model, even though you're carrying the full two-seventy-six billion in memory. That's the efficiency claim in one sentence — you pay the memory tax once, then generate cheap. The catch is that memory tax is still real: you're holding hundreds of gigs of weights whether or not you use them, which matters if you're serving many users at once and VRAM becomes the bottleneck instead of compute.

[21:49:00] David: Okay, compare that to Kimi.

[21:49:15] Eve: […] Kimi sits at the brutal end of that spectrum, and there's no small variant to soften it. Kimi K3.1 Pro is two-point-eight trillion total with a hundred and four billion active — the biggest open-weight model ever attempted — and the weights aren't even out yet, promised by November fourth. When they land, the floor is eight B300 or GB300 cards, about two-point-three terabytes of VRAM, and Moonshot's own production recommendation is a supernode of sixty-four-plus accelerators. There's no Inkling-Small equivalent, no twenty-seven-billion dense model you can grab on one GPU. The closest thing is the older K2 series at a trillion total and thirty-two billion active, which still needs roughly six hundred gigs at four-bit — eight H200s — though Unsloth's extreme one-point-eight-bit GGUF can squeeze it onto a single twenty-four-gig card with heavy RAM offload, at single-digit token speeds. So the intuitive picture is: Qwen gives you a ladder from laptop to datacenter, Inkling-Small is the clever middle rung, and Kimi is one rung only, and it's at the top of the building. The bet Moonshot is making is that being near-frontier on open weights matters more than being runnable — and with K3.1 Pro scoring fifty-five on Artificial Analysis, four points off Claude Opus, that bet is looking more credible than it did a year ago. One wrinkle worth knowing: they deliberately stripped cyber capabilities from the open weights, scoring eighty-one on their internal cyber bench versus ninety for the unfiltered checkpoint. So even the "open" version isn't the full model.

The thing I'd watch is whether a K3-class model eventually gets a distilled or sparse sibling the way Inkling did — because right now Kimi's open-weights story is all-or-nothing, and that limits who can actually use it. […]