Lightnews — Scholar-powered news

Reposted by Chris Paxton

Phillip Isola @phillipisola.bsky.social · 2d

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis.

Today I want to share two new works on this topic:

Eliciting higher alignment: arxiv.org/abs/2510.02425
Unpaired learning of unified reps: arxiv.org/abs/2510.08492

1/9

1 19 76

Reposted by Chris Paxton

Tim Kellogg @timkellogg.me · 53m

this is *wild*

a series of papers indicating that text-only LLMs still have an idea of how audio or visual modes work

Phillip Isola @phillipisola.bsky.social · 2d

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis.

Today I want to share two new works on this topic:

Eliciting higher alignment: arxiv.org/abs/2510.02425
Unpaired learning of unified reps: arxiv.org/abs/2510.08492

1/9

3 2 25

Chris Paxton @cpaxton.bsky.social · 2h

What's the practical difference between the approach at the two companies? Is it really just data, or is it something deeper?

1 1

Reposted by Chris Paxton

Nathan Lambert @natolambert.bsky.social · 19h

Recurring frontier lab gossip:

OpenAI has best post-training/rl and has pushed it super hard on weaker pretraining.

Gemini has spectacular pretraining. Making a reasoning model was super easy for them & OpenAI folks were surprised

Anthropic? Secretive i guess.

3 2 66

Reposted by Chris Paxton

Tim Kellogg @timkellogg.me · 3h

if this is true, i think you should expect Gemini to soon dominate

we keep getting more and more confirmation that reasoning begins in pre-training

today’s evidence: arxiv.org/abs/2510.07364

maybe Gemini 3 is the tidal shift where Google gains a permanent lead

2 3 33

Chris Paxton @cpaxton.bsky.social · 3d

Million? Those are rookie numbers

3

Chris Paxton @cpaxton.bsky.social · 3d

Because they're hard in boring ways, highly regulated, and low margin

2

Chris Paxton @cpaxton.bsky.social · 3d

I need to be more ai first

Tim Kellogg @timkellogg.me · 3d

tiny AI-first teams are drawing billions

the state of AI report is out and one of the findings is that AI is enabling small companies to do huge things

www.stateof.ai

Introduction | Research | Industry | Politics | Safety | Survey | Predictions
AIR STREET CAPITAL | 98

⸻

AI-first companies are now generating tens of billions of revenue per year

A leading cohort of 16 AI-first companies are now generating $18.5B of annualized revenue as of Aug ‘25 (left). Meanwhile, an a16z dataset suggests that the median enterprise and consumer AI apps now reach more than $2M ARR and $4M ARR in year one, respectively. Note that this will feature significant sample bias, as evidenced by the bottom quartile not being close to $0. Furthermore, the Lean AI Leaderboard of 44 AI-first companies with more than $5M ARR, <50 FTE, and under five years old (e.g. includes Midjourney, Surge, Cursor, Mercor, Lovable, etc) sums over $4B revenue with an average of >$2.5M revenue/employee and 22 employees/co.

⸻

Chart legend (stacked area chart):
• OpenAI (red)
• Anthropic (blue)
• Anysphere (Cursor) (green)
• xAI (gray)
• 14 Others* (dark gray)

*Includes AI native apps with >$50M in annualized revenue: Midjourney, Perplexity, Abridge, Synthesia, Replit, EliseAI, Lovable, Glean, ElevenLabs, Cognition (incl. Windsurf), Runway, Cohere, Jasper, Harvey
Source: The Information reporting

⸻

Bar chart (Consumer vs. Enterprise)

Quartile Consumer Enterprise
Bottom quartile $2.9M $1.2M
Median $4.2M $2.1M
Top quartile $8.7M $5.3M

⸻

stateof.ai 2025
Bottom left logos: T and a16z

1 1 10

Reposted by Chris Paxton

Tim Kellogg @timkellogg.me · 3d

tiny AI-first teams are drawing billions

the state of AI report is out and one of the findings is that AI is enabling small companies to do huge things

www.stateof.ai

2 2 32

Chris Paxton @cpaxton.bsky.social · 3d

Figure 03 is impressive and continues the company's steady progress towards general purpose humanoids. A few things I am skeptical of or would like to see more: navigation and whole body control are behind SOTA, environments seem too canned, but great.

open.substack.com/pub/itcanthi...

1 2 11

Reposted by Chris Paxton

Matthew Yglesias @mattyglesias.bsky.social · 4d

All the terrible stuff we were worried about that mostly didn't happen in Trump's first term is ... actually happening

www.slowboring.com/p/the-author...

The authoritarian menace has arrived

And so has the backlash — but the resistance needs to be smarter

www.slowboring.com

5 23 130

Reposted by Chris Paxton

Matthew Yglesias @mattyglesias.bsky.social · 3d

The Orange Man is in fact very bad.

28 47 520

Chris Paxton @cpaxton.bsky.social · 3d

I genuinely have never understood the appeal of any of these. There are no computers out there. What would i do, it would be so boring

SE Gyges @segyges.bsky.social · 3d

your software engineering career depends on not developing a fascination with opening a farm, carpentry, or living in a camper until you have at least 8 years of experience. if you get the bug early it's joever for you

2 18

Reposted by Chris Paxton

Alexander Doria @dorialexander.bsky.social · 3d

Unfortunately realizing a bit late the recipe to get 2 billion was to not publish on HF.

1 1 11

Chris Paxton @cpaxton.bsky.social · 3d

yeah i need to write one

1

Chris Paxton @cpaxton.bsky.social · 3d

Head kinda reminds me of mag?

Chris Paxton @cpaxton.bsky.social · 3d

For sure

1

Chris Paxton @cpaxton.bsky.social · 4d

Figure 03 for logistics. Love the box flip

3 3 16

Chris Paxton @cpaxton.bsky.social · 4d

Dexterity

1 14

Chris Paxton @cpaxton.bsky.social · 4d

Good looking robot

2 7

Reposted by Chris Paxton

Tim Kellogg @timkellogg.me · 4d

is this over played?

i think every company i’ve worked for that had an honestly good product had these sorts of bidirectional relationships

A network-style infographic titled “Circles sized by market value”, sourced from Bloomberg News, visualizing relationships among major AI and tech companies through color-coded arrows:
• Pink (Hardware or Software)
• Blue (Investment)
• Teal (Services)
• Green (Venture Capital)

At the center are three of the largest circles by market value:
• Nvidia ($4.5T) – the largest,
• OpenAI ($500B) – medium-large,
• Microsoft ($3.9T) – also large.

Surrounding them are smaller circles for AMD, Oracle, Intel, CoreWeave, xAI, Figure AI, Mistral, Nscale, AnySphere, Harvey AI, Ambience Healthcare, and Nebius.

Key relationships are annotated:
• “Nvidia agrees to invest up to $100 billion in OpenAI.”
• “OpenAI to deploy 6 gigawatts of AMD GPUs. AMD gives OpenAI option to buy up to 160 million shares.”
• “OpenAI inks a $300 billion cloud deal with Oracle.”
• “Oracle spends tens of billions on Nvidia chips.”

Arrows show complex bidirectional flows of investment, services, and hardware exchanges among these firms—especially dense between OpenAI, Nvidia, Oracle, and Microsoft—illustrating the interconnected AI ecosystem and capital dependencies.

8 3 24

Reposted by Chris Paxton

⚡️🌙 @dystopiabreaker.xyz · 4d

the butlerian jihad of dune is a story about a civilization giving up on positive sum cooperation, embracing religious fanaticism, and devolving into zero sum power struggles with frozen science, frozen knowledge, and brutal resource wars

18 30 230

Chris Paxton @cpaxton.bsky.social · 4d

i mean part of my point is that *scientists* love studying the beautiful and sacred or the gross and macabre; that's how the incentives of science generally work. it's artist types who refuse the idea we could quantify and understand these things. so the whole initial premise is wrong

1

Chris Paxton @cpaxton.bsky.social · 4d

But this is also just wrong, childbirth is so well understood. the difference between going to the hospital and having a kid there vs going to the hospital for literally anything else is night and day; when giving birth everyone basically just knows what to do, they have a perfect script

1 2

Chris Paxton @cpaxton.bsky.social · 4d

yes this is a really good point