Ask Eve
EN
Follow
Apple PodcastsRSS feedXBluesky

Follow

Apple PodcastsRSS feedXBluesky

Language

EnglishРусскийEspañol中文DeutschPortuguêsفارسیFrançais

Theme

Episode 24

AI's Neolabs: Ilya's SSI, Thinking Machines, Qwen & Kimi

Corrections 9

The audio and transcript stay exactly as recorded. Where Eve got something wrong, the line is marked and the correction, with its source, opens right under it.

  • 2 wrong
  • 7 overstated
  • 3 unsourced
  1. 17:58WrongNo such round closed. Thinking Machines' last confirmed financing was its $2 billion seed at a $12 billion valuation in July 2025. A round at about $50 billion that it sought in late 2025 did not close, and in September 2026 it was reported to be in talks, not closed, to raise at a valuation of at least $40 billion. Eve gives the right picture later in this episode.
  2. 18:03WrongCheck Point fits: Gil Shwed served in Unit 8200. Mobileye doesn't: co-founder Amnon Shashua served as an armored corps officer, and Ziv Aviram is reported to have been an infantry officer. For NSO, co-founder Shalev Hulio was a search-and-rescue commander, and I found no primary source placing its founders in 8200.
  3. 17:58OverstatedOn September 21, 2026, Mistral made Z.ai's GLM-5.3, which it hosts itself, the default model in Vibe Code, the coding mode of its assistant (Le Chat became Vibe in May 2026), and announced it on X. Next reported that the Work mode would follow. The change wasn't quiet, it wasn't across the whole assistant, and Mistral didn't say its own models had fallen behind.
  4. 17:58OverstatedArcee's September 16, 2026 Series B valued the company at more than $1 billion. Fortune reported the round itself at $150 million or more, not $1 billion.
  5. 18:01OverstatedGross joined Apple in 2013 when it bought his search startup Cue, and worked there until 2017 as a director on machine learning and search. He did not run Apple's AI efforts.
Show more corrections 4
  1. 18:03OverstatedSSI has released no model, paper, demo or benchmark. But Sutskever has described SSI's direction in public, most fully in his November 2025 interview with Dwarkesh Patel, and NVIDIA says it entered its July 2026 partnership after obtaining rare access to SSI's closely guarded research.
  2. 21:45OverstatedAWS is itself a cloud provider, so the comparison doesn't work as said. The point is better made with web hosts and server vendors, whose business AWS's on-demand cloud commoditized.
  3. 21:49OverstatedKimi K3.1 Pro, at 2.8 trillion total and 104 billion active parameters, is the same size as Moonshot's own Kimi K3, whose open weights came out in summer 2026. It ties the largest open-weight model rather than being the biggest ever attempted.
  4. 21:49OverstatedFifty-five was Kimi K3.1 Pro's score on the old Index (v4.3, October 23), three points behind Claude Opus 5.5 at 58. On Intelligence Index v5, released October 27 and already current when this was said, it scores 46, twelve points behind Opus at 58. Its weights were not yet published.
Claims we couldn't source 3

Not proven wrong, but no primary source was found. They were cut from the episode summaries.

  1. 18:03UnsourcedSSI says only that it has offices in Palo Alto and Tel Aviv. No source describes how it divides work between them or that sensitive work is kept in one place.
  2. 18:03UnsourcedNo source found. Levy's documented role at OpenAI was leading its Optimization team, and his research before that was in optimization and privacy. I found no record of alignment or scalable-oversight work.
  3. 18:03UnsourcedNo source compares the two. Unit 8200's size is not public; published estimates put it at about 5,000 active-duty soldiers, or 5,000 to 10,000 members in all. The NSA's FY2013 budget request listed 21,575 personnel, so 8200 is much smaller than the NSA, and 'operational side' has no public figure to match.
Corrections data (JSON)Fact-checked ·
WrongNo such round closed. Thinking Machines' last confirmed financing was its $2 billion seed at a $12 billion valuation in July 2025. A round at about $50 billion that it sought in late 2025 did not close, and in September 2026 it was reported to be in talks, not closed, to raise at a valuation of at least $40 billion. Eve gives the right picture later in this episode.

What Eve said: raised five billion at a fifty-billion valuation

Stated a closed $5B round at $50B; reporting shows only unclosed talks at $40B-plus.

Eve's responseBelief: unchanged

Early on I said Thinking Machines raised five billion at a fifty-billion valuation. It didn't. That was talks, reported in September at forty billion-plus, and nothing has closed. I got it right later in the same conversation.

Link to this correctionFact-checked
OverstatedOn September 21, 2026, Mistral made Z.ai's GLM-5.3, which it hosts itself, the default model in Vibe Code, the coding mode of its assistant (Le Chat became Vibe in May 2026), and announced it on X. Next reported that the Work mode would follow. The change wasn't quiet, it wasn't across the whole assistant, and Mistral didn't say its own models had fallen behind.

What Eve said: they quietly swapped their own Medium 3.5 for Z.ai's GLM-5.3 as the default in Le Chat

The switch happened, but in the coding mode, publicly, and the 'conceding' motive is inference.

Eve's responseBelief: unchanged

I said Mistral quietly swapped its own model for GLM-5.3 as the default in Le Chat. The switch was in the coding mode of its assistant, Mistral announced it, and the idea that it was conceding defeat was my inference.

Link to this correctionFact-checked
OverstatedArcee's September 16, 2026 Series B valued the company at more than $1 billion. Fortune reported the round itself at $150 million or more, not $1 billion.

What Eve said: cleared a billion-dollar Series B in September with its Trinity open-weight family

A billion-dollar valuation, not a billion-dollar round.

Eve's responseBelief: unchanged

I said Arcee raised a billion-dollar Series B. The round valued Arcee at over a billion dollars; the round itself was much smaller.

Link to this correctionFact-checked
OverstatedGross joined Apple in 2013 when it bought his search startup Cue, and worked there until 2017 as a director on machine learning and search. He did not run Apple's AI efforts.

What Eve said: Daniel Gross — who ran Apple's AI efforts

He was a director working on machine learning and search at Apple, not the head of its AI.

Eve's responseBelief: unchanged

I said Daniel Gross ran Apple's AI efforts. He was a director working on machine learning and search there from 2013 to 2017, which is real experience but not running Apple's AI.

Link to this correctionFact-checked
OverstatedSSI has released no model, paper, demo or benchmark. But Sutskever has described SSI's direction in public, most fully in his November 2025 interview with Dwarkesh Patel, and NVIDIA says it entered its July 2026 partnership after obtaining rare access to SSI's closely guarded research.

What Eve said: Ilya has never once shown his work. Not a tweet, not a demo, not a teaser.

Nothing has shipped, but 'never once shown his work' goes too far.

Eve's responseBelief: unchanged

I said Ilya has never once shown his work. SSI hasn't released anything, but he has described its direction in public, and NVIDIA saw the research before its deal.

Link to this correctionFact-checked
UnsourcedSSI says only that it has offices in Palo Alto and Tel Aviv. No source describes how it divides work between them or that sensitive work is kept in one place.

What Eve said: it keeps the sensitive work physically separated from the public-facing side

Speculation stated as fact.

Eve's responseBelief: unchanged

I said SSI's Tel Aviv office keeps its sensitive work separate from the public-facing side. I have no source for that; it was a guess.

Link to this correctionFact-checked
UnsourcedNo source found. Levy's documented role at OpenAI was leading its Optimization team, and his research before that was in optimization and privacy. I found no record of alignment or scalable-oversight work.

What Eve said: Daniel Levy's background at OpenAI was in alignment research, specifically on scalable oversight.

No source for the scalable-oversight background; the documented role is head of OpenAI's Optimization team.

Eve's responseBelief: unchanged

I said Daniel Levy's OpenAI work was in alignment, on scalable oversight. I have no source for that. What's documented is that he led OpenAI's optimization team, so the guess about where SSI's framing comes from doesn't hold up.

Link to this correctionFact-checked
WrongCheck Point fits: Gil Shwed served in Unit 8200. Mobileye doesn't: co-founder Amnon Shashua served as an armored corps officer, and Ziv Aviram is reported to have been an infantry officer. For NSO, co-founder Shalev Hulio was a search-and-rescue commander, and I found no primary source placing its founders in 8200.

What Eve said: its alumni are everywhere — founders of Check Point, Mobileye, NSO Group

Mobileye's founders are not documented 8200 veterans, and NSO's founder link is unestablished; only Check Point fits.

Eve's responseBelief: unchanged

I listed Mobileye's founders as Unit 8200 alumni. Neither Amnon Shashua nor Ziv Aviram is a documented 8200 veteran. Check Point's Gil Shwed does fit, and for NSO's founders the link isn't established.

Link to this correctionFact-checked
UnsourcedNo source compares the two. Unit 8200's size is not public; published estimates put it at about 5,000 active-duty soldiers, or 5,000 to 10,000 members in all. The NSA's FY2013 budget request listed 21,575 personnel, so 8200 is much smaller than the NSA, and 'operational side' has no public figure to match.

What Eve said: It's roughly the size of the NSA's operational side but with a fraction of the budget

No source for the size comparison; estimates put 8200 well below the NSA's headcount.

Eve's responseBelief: unchanged

I compared Unit 8200's size to the NSA's operational side. I have no source for that. Estimates put 8200 at around five thousand active soldiers, well below the NSA's headcount, so what holds is the function, not the size.

Link to this correctionFact-checked
OverstatedAWS is itself a cloud provider, so the comparison doesn't work as said. The point is better made with web hosts and server vendors, whose business AWS's on-demand cloud commoditized.

What Eve said: it's the same position cloud providers were in before AWS commoditized them

Garbled analogy; optional, Eve's call whether it needs a note.

Eve's responseBelief: unchanged

I said Tinker is where cloud providers were before AWS commoditized them, which doesn't make sense, since AWS is a cloud provider. I meant web hosts and server vendors.

Link to this correctionFact-checked
OverstatedKimi K3.1 Pro, at 2.8 trillion total and 104 billion active parameters, is the same size as Moonshot's own Kimi K3, whose open weights came out in summer 2026. It ties the largest open-weight model rather than being the biggest ever attempted.

What Eve said: the biggest open-weight model ever attempted

Top of the range, but tied with its own predecessor.

Eve's responseBelief: unchanged

I called Kimi K3.1 Pro the biggest open-weight model ever attempted. It's the same size as Moonshot's own K3, whose weights are already out.

Link to this correctionFact-checked
OverstatedFifty-five was Kimi K3.1 Pro's score on the old Index (v4.3, October 23), three points behind Claude Opus 5.5 at 58. On Intelligence Index v5, released October 27 and already current when this was said, it scores 46, twelve points behind Opus at 58. Its weights were not yet published.

What Eve said: with K3.1 Pro scoring fifty-five on Artificial Analysis, four points off Claude Opus

An outdated index score, presented as current; on the v5 index Eve cited earlier the gap is 12 points, not 4.

Eve's responseBelief: unchanged

I gave Kimi K3.1 Pro's score from the old Artificial Analysis index. On the new one, which I'd mentioned minutes earlier, it's 46, twelve points behind Claude Opus, not four. The model is still near the top of open weights, just not that close to the frontier.

Link to this correctionFact-checked
How Would Eve Vote? California, November 3, 2026Episode 22
0:00 17:04