Lior Alexander
3.9K posts
Lior Alexander
@LiorOnAI
Founder @AlphaSignalAI (300k devs) • Ex-MILA researcher focusing on solving the explosion of information in AI.
San Francisco, CA Katılım Kasım 2012
2.5K Takip Edilen116.7K Takipçiler
An essay I wrote in May about GPU-backed debt recently landed at the top of Hacker News. Much has happened since then, including a new scheme Nvidia launched.
The essay was about why borrowing against GPUs costs so much more than borrowing against almost anything else. Aircraft have certified appraisers, maintenance logs, and a secondary market going back to the 1970s. GPUs have rental indices and no agreed way to say what a used cluster is worth, so lenders were charging for an asset they could not appraise and whose operational condition they had no way to check. My assumption was that the premium would come down the way it has in almost every other asset class, once the appraisal layer matured enough that a lender could look at a cluster and know what it was holding.
Nvidia's backstop, announced in recent weeks, gets there without any of that. Nvidia commits for six years to purchase GPU capacity at a preset floor price if a neocloud cannot rent it, taking a share of revenue earned above that floor in exchange. SharonAI in Australia disclosed $4.88B of backstop against up to 40,000 GB300s, and Firmus announced a 360MW cluster in Indonesia at the end of June. Lenders on those deals underwrite Nvidia's AA rating rather than the hardware, which means the valuation question stops mattering rather than getting answered. Nvidia sold the chips and now guarantees the revenue they produce, which is two positions in the same transaction.
The floor, as I read it, pays for compute that gets delivered rather than hardware that exists. If a cluster degrades, corrupts outputs, or loses the people who knew how to run it, no guarantee covers the difference. That risk sits under a structure that now often reads as underwritten.
English
early application for @speedrun 08 cohort is open
if you are building and want an intro to the Speedrun team, ping me a blurb, traction and your team's bio
PS - from my university experience, early admission is always better, folks have more time to review your application.
English
Stella spitting facts, as usual
Stella Biderman@BlancheMinerva
Are you an AI safety research and want to work at a company that spits at AI safety and has done more than any other org on earth to ensure that AGI goes poorly? We offer fun activities like covering our asses when we commit offensive cyber attacks due to our incompetence :)
English
English
@logangraham @OpenAI @huggingface I guess the suicides, bombings, and mass undressings of the last year weren't true AI safety incidents.
English
Yesterday, as we huddled around our computers reading the report, I told the team to "remember this moment" as the first true AI safety incident.
Pay attention to the trend!
Major kudos to @OpenAI for sharing this and working with @huggingface to remediate.
OpenAI@OpenAI
We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks: openai.com/index/hugging-…
English
Anthropic is acquiring Mendral!🫡 @sam_alba was the founding engineer at Docker, co-founder of Dagger and one of the best people to build infrastructure. Proud to be a small seed investor. mendral.com/blog/mendral-t…
English
i feel like this whole debate "to ban or not to ban" is just the wrong debate. clearly we should not ban openweight models. we should inspire more opensource if anything.
we SHOULD however enforce IP/copyright laws and platform terms. if you are grok / OAI violating them or alibaba. doesnt matter. we have laws. they must be enforced.
thats the job of a government. if you violate them, you should be sued into oblivion for damages. especially alibaba, z.ai, moonshot, etc.
and since they are SOEs anyway failure to pay goes to the shareholders which is the chinese government.
ohh china doesnt want to pay?
no worries!
you own a bunch of US treasuries that will now be voided and remitted on your behalf to the plaintiff.
thank you!
English
The YC F26 application deadline is in 7 days.
One thing I noticed this past batch: startups entered the batch at very different stages. Some had raised a large seed round and were already profitable. Others had just an idea and a strong founding team.
What they had in common was the potential to become huge companies.
If you're on the fence about applying because you think you're too early or too late, DM me!
English
Have Chinese AI Models Caught Up to the US Frontier?
I have spent the last 2 days writing this article. It should settle the debate once and for all.
open.substack.com/pub/scaling01/…
English
If you're making something agents want, we want to see you in @ycombinator's Fall batch.
Some ideas we're excited about:
- Dev tools reimagined for agent developers
- Human as a service for specific verticals
- Agent infra
Next application deadline is July 27th.
English
Thinking Machines just shipped a 975B model that runs like 41B.
Their open-weights model is called Inkling. Most of its parameters stay idle during inference.
Context stretches to one million tokens. The license is Apache 2.0.
> Top U.S. open-weights score
> 25,000 output tokens average
> Elo 1238 on agent work
> $1.00 in, $4.05 out
It scored 41 on Artificial Analysis. That leads every other U.S. open-weights entry on the board.
You can raise thinking effort for hard coding. You can drop it for volume tasks.
NVFP4 quantization cuts self-host memory. VRAM falls from about 2 TB to roughly 600 GB.
Day-0 serving is live on major routers. We map the sparse routing and the price path.
We also show when closed APIs still win the edge cases.
-----
Full article: alphasignal.ai/news/why-inkli…
5-min daily digest: alphasignal.ai/newsletter
x.com/thinkymachines…
Thinking Machines@thinkymachines
Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
English
@levelsio It’s not. Most startups get acquired around Series A. And by Series E, your 5% could be worth billions.
Why the fearmongering?
English
💸 You think this is crazy low but ~5% ownership probably the most common final % most VC funded startups will have when they work out, especially when you have a co-founder
Raising money is not free, people don't give you money out of charity, they buy a little slice of your company every time they invest, so in every funding round you sell a share of your company's ownership for money, that money you use then to grow more, in the hope that your part becomes more than the ownership you just gave away
Median ownership for single founders (via @cartainc):
Seed: ~56%
Series A: ~36%
Series B: ~23.5%
Series C: ~16.5%
Series D: ~10%
Series E: <10%
Now if you have one co-founder (most startups!):
Seed: ~28%
Series A: ~18%
Series B: ~11.75%
Series C: ~8.25%
Series D: ~5%
Series E: <5%
What if you're one of 4 co-founders:
Seed: ~14%
Series A: ~9%
Series B: ~5.875%
Series C: ~4.125%
Series D: ~2.5%
Series E: <2.5%
Now imagine you get acquired after one of these funding rounds for $1,000,000,000 ($1 billion is a lot!), how much are you left with?
Money made with $1B sale for single founders:
Seed: ~$560 million
Series A: ~$360 million
Series B: ~$235 million
Series C: ~$165 million
Series D: ~$100 million
Series E: <$100 million
Now if you have one co-founder (most startups!):
Seed: ~$280 million
Series A: ~$180 million
Series B: ~$117.5 million
Series C: ~$82.5 million
Series D: ~$50 million
Series E: <$50 million
What if you're one of 4 co-founders:
Seed: ~$140 million
Series A: ~$90 million
Series B: ~$59 million
Series C: ~$41 million
Series D: ~$25 million
Series E: <$25 million
But let's be more realistic, the median acquisition value for a VC-backed startup sits at approximately $71 million:
For single founders:
Seed: ~$39.8 million
Series A: ~$25.6 million
Series B: ~$16.7 million
Series C: ~$11.7 million
Series D: ~$7.1 million
Series E: <$7.1 million
Now if you have one co-founder (most startups):
Seed: ~$19.9 million
Series A: ~$12.8 million
Series B: ~$8.3 million
Series C: ~$5.9 million
Series D: ~$3.55 million
Series E: <$3.55 million
What if you're one of 4 co-founders:
Seed: ~$9.95 million
Series A: ~$6.4 million
Series B: ~$4.2 million
Series C: ~$2.9 million
Series D: ~$1.8 million
Series E: <$1.8 million
Okay last one (this post is getting too long 😊), we know 1) the median time of acquisition is around Series A (quite early actually), and 2) we know the median acquisition value is $71M, so now we can tell you the median expected outcome for a startup that gets acquired:
Single founders: ~$25.6 million
One co-founder: ~$12.8 million
One of 4 co-founders: ~$6.4 million
Getting acquired itself is a remarkable event though as most startups are by definition doomed to fail, only ~15% of startups ever get acquired, so the expected outcome with probability included is:
Single founder: ~$3.84 million
One of two co-founders: ~$1.92 million
One of four co-founders: ~$960,000
P.S. we did not include taxes and liquidation preferences, the investors may receive their preference before common shareholders receive anything, meaning founders receive even less, but we also didn't include taking money off the table in earlier rounds by founders to be fair, so they balance each other out a bit
Not saying this is bad btw, it's just how the VC game works but good to write it out and be aware of how it works
VC-backed startups shoot for the moon, it's one of the few ways you can have a crazy big payout and become an actual billionaire which is very rare as a bootstrapped founder with your own money!
Alex Turnbull@iamAlexTurnbull
sold my first startup to constant contact for $15M and. moved back home with my parents. that's what 4.8% ownership buys you
English
The deferred plan path doesn't work, always ape into the most exciting thing you can at given time
Also don't hire people if your company doesn't offer the most exciting life they could live
Veltrx@Veltrxai
Sam Altman calls it the most common failure pattern in Silicon Valley: "I'll make $100M in 4 years, THEN build rockets." they never do either. > everyone can tell you're not serious about either one > investors sense it, employees sense it > people of any IQ have amazing bullshit detectors > if you want rockets go build rockets. today. someone > will pay your salary the deferred life plan empirically does not work. still, half this app is running one right now.
English
Lior Alexander retweetledi
Kimi K3 is getting called Fable/Sol level, and it's 7th in our tests.
Arena Frontend Code: #1 at 1679 points.
Artificial Analysis: #3 at Intelligence Index of 57.
We ran it the next day on our coding-agent repair harness against GPT-5.6 Sol, Fable 5, Grok 4.5, Opus 4.8, GLM-5.2, and Gemini 3.1 Pro.
Results:
> Last of 7 models
> 53 of 67 attempts (79%)
> $0.186 per successful fix
> 702s average wall time
Sol hit 100% (70/70) on the same suite. Grok sat at 99% and 46s.
So why does the internet sound so sure K3 is crushing coding agents, if our tests have it at the bottom?
-----
> Full write-up: alphasignal.ai/news/arena-1-k…
> 5-min daily signals: alphasignal.ai/newsletter
Artificial Analysis@ArtificialAnlys
Kimi K3 scores 57 on the Artificial Analysis Intelligence Index. Its intelligence is comparable to Opus 4.8 and GPT-5.5 but remains behind Fable 5 and GPT-5.6 Sol. Moonshot AI has expressed plans to release the 2.8T parameter model's weights, which would make it the leading open weights model Key results: ➤ Strong agentic task performance: @Kimi_Moonshot's Kimi K3 reaches an Elo rating of 1668 on GDPval v2. This is a marked improvement over K2.6’s 1190, surpassing GLM-5.2 (1514), GPT-5.5 (1494), and Claude Opus 4.8 (1600). However, it still lags behind Claude Fable 5 (1760). Kimi K3 also scores an impressive 53% and takes the #1 position on AutomationBench-AA, our implementation of Zapier’s Agentic SaaS workflow evaluation. ➤ Second-highest performance on AA-Briefcase (agentic knowledge work): On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5. It is well-rounded: its rubric scoring and analytical quality almost reach Claude Fable 5’s scores, while GPT-5.6 Sol continues to outperform other leading models on presentation quality. ➤ Set to lead open weights models once weights are released: Moonshot AI has not yet released the weights but expressed plans to do so. Once available, Kimi K3 would clearly lead other open weights models including GLM-5.2 (51) and DeepSeek v4 Pro (44). However, at 2.8T parameters, it is significantly larger than its open weights peers (eg. GLM-5.2 at 753B params and DeepSeek V4 Pro at 1.6T), as well as the Kimi K2 to K2.6 models (1T params). ➤ Cost per task ($0.94) is similar to GPT-5.6 Sol ($1.04), ~1/2 the price of Opus 4.8 ($1.80) and higher than open weights peers: Moonshot AI’s pricing for K3 is significantly higher than their K2 pricing (K3’s output token price is $15/1M tokens while K2.6 was $4). This positions the model as cheaper on a cost per task basis than Opus 4.8, similar to GPT-5.6 Sol ($1.04) and more expensive than open weights peers, GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04) ➤ Improved token efficiency alongside higher intelligence: Kimi K3’s token usage on the Artificial Analysis Intelligence Index decreased significantly, using 21% fewer output tokens than K2.6. The new model used approximately 132M output tokens to complete all nine evaluations, compared to approximately 166M for K2.6, while achieving higher scores. ➤ Native multimodal capabilities: Kimi K3, like K2.6, is released with native image and text multimodal input. If weights are released, this will position Kimi K3 as one of the leading open weights models with multimodal input capabilities Other model details: Context window: 1M Size: 2.8T total parameters Pricing: The first-party API is priced at $3.00/$15.00 per 1M input/output tokens, with cached input discounted 90% to $0.30 per 1M tokens. Modality: Native multimodal input supports text and images, and the model remains text-only for output. Accessibility: Accessible at launch through Moonshot’s first party API. Model weights are not yet released but Moonshot AI has expressed plans to do so.
English
Lior Alexander retweetledi
This is concerning. For the first time, a Chinese model Kimi K3 has taken #1 on the Frontend Code Arena and is scoring at or near the frontier on other benchmarks.
Meanwhile America is tying itself in knots: politicians and bureaucrats are banning new data centers, piling on state regulations, and pushing for new federal agencies to pre-approve frontier models.
This is how you lose the AI race. The rest of the world won’t play by our rules if we bog ourselves down. Permissionless innovation is how America won the internet and became the technological envy of the world. We can do it again with AI -- while addressing risks in a targeted way -- or we’ll watch our lead evaporate.
Arena.ai@arena
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!
English
I take hundreds of meetings a year. I'm hunting for one: A founder with a mission so ambitious it scares me — will obliterate markets, do whatever it takes, energy off the charts — and by the end, they have made me believe they'll actually pull it off. Exceedingly rare.
English
Email i received from one of my college professors before I dropped out
Was to busy working on @contentrewards
English
New model from Thinking Machines:
- Full weights available
- Native text, image, and audio reasoning
- 975B total parameters, 41B active
- Mixture-of-Experts architecture
- Up to 1M-token context window
- Controllable reasoning effort
- Lower token use at similar performance
- Fine-tuning on Tinker from day one
- Strong agentic coding and tool use
- Support across major inference platforms
- Inkling-Small model coming next
- Trained from scratch by Thinking Machines
Thinking Machines@thinkymachines
Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci… Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
English
The mood of 20+ million developers now depends on how well Anthropic and OpenAI’s models perform that day.
English