Bloomberg reports that Mirendil, founded by ex-Anthropic researchers to build recursive self-improving models, is in talks to raise up to $1B at a $5B valuation led by Kleiner Perkins with a13 participating, which would place frontier-scale capital behind an explicitly self-improvement-focused neo-lab launching a frontier model by early 2027.

state: seedheat: lowuncertainty: mediumconvergesscott: mediumai-research frontier-labs ai-fundingMirendilKleiner PerkinsAndreessen Horowitz

What is this?

Mirendil is a San Francisco AI lab founded in 2025 by ex-Anthropic researchers Behnam Neyshabur and Harsh Mehta, with ~20 staff drawn from Anthropic, xAI, DeepMind and OpenAI, explicitly targeting recursive self-improvement — infrastructure that lets scientists autonomously build and iterate their own custom AI models ('AI for AI for science'). It raised a $200M seed at a $1B valuation (June 2026, co-led by a16z and Kleiner Perkins with NVIDIA) and later signed a $100M+ multi-year Google Cloud compute deal. The new case — a reported up-to-$1B round at a $5B valuation — is partially corroborated: a Yahoo Finance headline via Distill confirms talks 'at $5 billion valuation', but the supplied snippets do not confirm the round size, the lead, or an early-2027 frontier-model timeline; those specifics rest on the Bloomberg report alone.

Why it matters to Scott

Converges: frontier-scale capital (Kleiner/a13, $5B) is now explicitly underwriting recursive self-improvement as the founding thesis — the same self-improving-loop architecture Scott has built frameworks around and argued needs structural control ('can't beats shouldn't'), so the world's money just arrived where his safety doctrine and neo-lab/capability-symmetry economics already sit. The radar tracks the recursive-self-improvement, frontier-labs and ai-financing concepts broadly but has no Mirendil actor page, so this is a new episode under existing concepts rather than a repetition — a dated-receipts moment for his safety and loop-compounding arguments, though as a financing report (not a technical artifact) it sharpens his claims more than it changes what he builds.

ip:concept.self-improving-loopsip:concept.capability-symmetryip:framework.architecture-not-vibesip:source.ai-doesnt-fear-deathradar:concept.recursive-self-improvementradar:concept.frontier-labsradar:concept.ai-financingradar:concept.compute-financing

queries asked of Scott's wikis

recursive self-improvement safety position arguments
neo-labs ex-frontier-lab startup wave thesis
agent harnesses for autonomous research / self-improving coding loops
frontier model compute costs vs seed-stage capital economics
democratized model training: open tools for building custom in-house models
AI safety critique of models rewriting their own code

Measured heat

now 0 pts/hpeak 0 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 482h

points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

09-21 14:00	⭐ origin echo-reconstructed	"Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to raise a new round of funding at a Natasha Mascarenhas and Shirin Ghaffary (Bloomberg) on blog (echo) · attributed from hn.story.49806453	—
09-22 19:04	first on hacker news · published · +29.1h	Mirendil in talks to raise $1B at $5B valuation led by Kleiner Perkins utiiiD	—
09-22 19:04	amplified on hacker news 👑	hn.story.49806453 utiiiD	peak 2 · 0 comments · 98% of case engagement
09-22 19:20	our radar first saw it · +29.4h	discovery anchor: hn.story.49806453	—

pace: p9 vs 1032 stories at the 336h mark (now 482h old) — behind addom-local-coding-harness (0.5x)

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Mirendil in talks to raise $1B at $5B valuation led by Kleiner Perkins Retrieved article excerpt Open article · Retrieved 2026-09-22T19:23:36.576667+00:00 [Bloomberg](https://www.bloomberg.com/company "Bloomberg") # Ex-Anthropic Staffers’ Self-Improving AI Startup in Talks to Raise at $5 Billion Value An attendee at Anthropic PBC's Code with Claude developer conference in London. · Bloomberg · Bloomberg Natasha Mascarenhas and Shirin Ghaffary Wed, September 23, 2026 at 2:16 AM GMT+10 3 min read (Bloomberg) -- Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to raise a new round of funding at a $5 billion valuation, including the investment, according to people familiar with the matter.Kleiner Perkins is in talks to lead the current financing effort, which could bring up to $1 billion in new capital to the startup, according to the people, who spoke on the condition of anonymity due to discussing sensitive information. Andreessen Horowitz is in discussions to invest in the round as well, the people said. Most Read from Bloomberg - [Paramount to Settle Lawsuits, Paving Way for Warner Bros.](https://www.bloomberg.com/news/articles/2026-09-21/paramount-to-settle-lawsuits-paving-way-for-warner-bros-merger?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) - [Nvidia's Stock Is Flashing a Warning Sign as Valuation Falls](https://www.bloomberg.com/news/articles/2026-09-22/nvidia-s-stock-is-flashing-a-warning-sign-as-valuation-falls?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) - [Qatar Energy Chief Says Bessent 'Wrong' About Hormuz Future](https://www.bloomberg.com/news/articles/2026-09-20/qatar-energy-chief-says-bessent-wrong-about-hormuz-s-future?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) - [Stocks Jump on AI Optimism as AMD Tops $1 Trillion: Markets Wrap](https://www.bloomberg.com/news/articles/2026-09-20/us-stock-futures-up-ahead-of-talks-dollar-steady-markets-wrap?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) - [AI Risk Is Everywhere and It's Making Billion Dollar Funds Nervous](https://www.bloomberg.com/news/features/2026-09-20/ai-boom-is-making-diversifying-investments-tough-for-wall-street?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) The discussions come three months after it was announced that Kleiner and Andreessen led a $200 million seed round for the young startup. That deal valued the business at $1 billion in its first round of financing. Mirendil, now with more than 20 staffers, was started earlier this year with the goal of building widely accessible models capable of improving themselves with little to no help from humans. Some firms see this strategy, also known as "recursive self-improvement," as the future of AI development, but there have also been concerns about the security threats this form of AI could pose if it's achieved. The new funding would be used to help the startup launch its own model and train future ones. It's planning to launch a frontier model to support engineering and research work by the beginning of next year, according to two of the people.Mirendil declined to comment for the story. The deal talks suggest continued investor demand for neo-labs, or companies that are prioritizing ambitious research projects over short-term commercialization. A number of neo-labs have been on the fundraising circuit, including Thinking Machines Lab, which was started by OpenAI's former chief technology officer, and Periodic Labs, a scientific discovery company. There's also growing interest from software startups looking to build their own models in-house, driven by progress in open-source AI's capabilities. "Today, any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab – a process that is expensive and requires expertise concentrated in a small number of labs," Mirendil's website reads. " Our goal is to democratize frontier AI R&D and make it widely accessible."Recursive self-improvement is a massive area of interest for the leading AI developers, including Anthropic and OpenAI, because of its potential to speed the pace of innovation. For example, Anthropic recently shared that its Claude chatbot drives more than a quarter of the company's research and development work for AI. OpenAI, meanwhile, extensively uses its coding technology, Codex, to automate its own software and model development. Story Continues In recent weeks, however, some inside and outside the industry have warned that AI could someday extend beyond human control and create serious security risks. These fears are prompting calls to slow the pace of development — or at least create global standards for safely developing the tech. Mirendil has told stakeholders that a big part of its goal is to get advanced AI in the hands of more people, while solving the technical challenges of implementing strong safeguards. Training and running models is a costly endeavor, so Mirendil, which has only a fraction of the capital and resources that OpenAI and Anthropic have, is working on a cheaper, more automated way to build these models. Last month, the company announced a partnership with Google that included using Google Cloud TPUs, or tensor-processing units, to scale its model building efforts. The company also works with Nvidia Corp.'s computing infrastructure, and counts the chipmaker as an investor. Most Read from Bloomberg Businessweek - [People Hooked on Vapes Try a New Way to Quit: Cigarettes](https://www.bloomberg.com/news/articles/2026-09-18/to-quit-vaping-some-are-starting-to-smoke?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS) - [Hating on Polyester Is Back in Fashion](https://www.bloomberg.com/news/articles/2026-09-18/why-shoppers-are-turning-away-from-polyester-clothing?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS) - ['Death Sentences': Crafting America's Favorite Countertops Is Killing Workers](https://www.bloomberg.com/news/features/2026-09-14/silicosis-lawsuits-mount-as-workers-seek-billions-in-damages?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS) - [Trump's Quest to Be Consequential Puts the Entire World at Risk](https://www.bloomberg.com/news/articles/2026-09-18/iran-war-tariffs-and-trump-s-second-term-quest-to-be-consequential?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS) - [Fender Is Making Enemies With a Messy Fight Over Its Iconic Strat](https://www.bloomberg.com/news/features/2026-09-17/fender-takes-on-the-electric-guitar-industry-over-stratocaster-copyright?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS) ©2026 Bloomberg L.P. [Terms](https://guce.yahoo.com/terms?locale=en-US) and [Privacy Policy](https://guce.yahoo.com/privacy-policy?locale=en-US) [Privacy Dashboard](https://guce.yahoo.com/privacy-dashboard?locale=en-US) [More Info](https://finance.yahoo.com/more-info)	utiiiD	2	0
🟧 echo.blog ⭐	"Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to raise a new round of funding at a	Natasha Mascarenhas and Shirin Ghaffary (Bloomberg)	—	—

source

object

author

score

comments

🟧 hn

Mirendil in talks to raise $1B at $5B valuation led by Kleiner Perkins

Retrieved article excerpt

Open article · Retrieved 2026-09-22T19:23:36.576667+00:00

[Bloomberg](https://www.bloomberg.com/company "Bloomberg")

# Ex-Anthropic Staffers’ Self-Improving AI Startup in Talks to Raise at $5 Billion Value

An attendee at Anthropic PBC's Code with Claude developer conference in London. · Bloomberg · Bloomberg

Natasha Mascarenhas and Shirin Ghaffary

Wed, September 23, 2026 at 2:16 AM GMT+10 3 min read

(Bloomberg) -- Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to raise a new round of funding at a $5 billion valuation, including the investment, according to people familiar with the matter.Kleiner Perkins is in talks to lead the current financing effort, which could bring up to $1 billion in new capital to the startup, according to the people, who spoke on the condition of anonymity due to discussing sensitive information. Andreessen Horowitz is in discussions to invest in the round as well, the people said.

Most Read from Bloomberg

- [Paramount to Settle Lawsuits, Paving Way for Warner Bros.](https://www.bloomberg.com/news/articles/2026-09-21/paramount-to-settle-lawsuits-paving-way-for-warner-bros-merger?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)
- [Nvidia's Stock Is Flashing a Warning Sign as Valuation Falls](https://www.bloomberg.com/news/articles/2026-09-22/nvidia-s-stock-is-flashing-a-warning-sign-as-valuation-falls?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)
- [Qatar Energy Chief Says Bessent 'Wrong' About Hormuz Future](https://www.bloomberg.com/news/articles/2026-09-20/qatar-energy-chief-says-bessent-wrong-about-hormuz-s-future?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)
- [Stocks Jump on AI Optimism as AMD Tops $1 Trillion: Markets Wrap](https://www.bloomberg.com/news/articles/2026-09-20/us-stock-futures-up-ahead-of-talks-dollar-steady-markets-wrap?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)
- [AI Risk Is Everywhere and It's Making Billion Dollar Funds Nervous](https://www.bloomberg.com/news/features/2026-09-20/ai-boom-is-making-diversifying-investments-tough-for-wall-street?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)

The discussions come three months after it was announced that Kleiner and Andreessen led a $200 million seed round for the young startup. That deal valued the business at $1 billion in its first round of financing.

Mirendil, now with more than 20 staffers, was started earlier this year with the goal of building widely accessible models capable of improving themselves with little to no help from humans. Some firms see this strategy, also known as "recursive self-improvement," as the future of AI development, but there have also been concerns about the security threats this form of AI could pose if it's achieved.

The new funding would be used to help the startup launch its own model and train future ones. It's planning to launch a frontier model to support engineering and research work by the beginning of next year, according to two of the people.Mirendil declined to comment for the story.

The deal talks suggest continued investor demand for neo-labs, or companies that are prioritizing ambitious research projects over short-term commercialization. A number of neo-labs have been on the fundraising circuit, including Thinking Machines Lab, which was started by OpenAI's former chief technology officer, and Periodic Labs, a scientific discovery company. There's also growing interest from software startups looking to build their own models in-house, driven by progress in open-source AI's capabilities.

"Today, any lab trying to use AI in drug discovery, chemistry, biology, or robotics must also become a frontier AI lab – a process that is expensive and requires expertise concentrated in a small number of labs," Mirendil's website reads. " Our goal is to democratize frontier AI R&D and make it widely accessible."Recursive self-improvement is a massive area of interest for the leading AI developers, including Anthropic and OpenAI, because of its potential to speed the pace of innovation. For example, Anthropic recently shared that its Claude chatbot drives more than a quarter of the company's research and development work for AI. OpenAI, meanwhile, extensively uses its coding technology, Codex, to automate its own software and model development.

Story Continues

In recent weeks, however, some inside and outside the industry have warned that AI could someday extend beyond human control and create serious security risks. These fears are prompting calls to slow the pace of development — or at least create global standards for safely developing the tech. Mirendil has told stakeholders that a big part of its goal is to get advanced AI in the hands of more people, while solving the technical challenges of implementing strong safeguards. Training and running models is a costly endeavor, so Mirendil, which has only a fraction of the capital and resources that OpenAI and Anthropic have, is working on a cheaper, more automated way to build these models. Last month, the company announced a partnership with Google that included using Google Cloud TPUs, or tensor-processing units, to scale its model building efforts. The company also works with Nvidia Corp.'s computing infrastructure, and counts the chipmaker as an investor.

Most Read from Bloomberg Businessweek

- [People Hooked on Vapes Try a New Way to Quit: Cigarettes](https://www.bloomberg.com/news/articles/2026-09-18/to-quit-vaping-some-are-starting-to-smoke?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS)
- [Hating on Polyester Is Back in Fashion](https://www.bloomberg.com/news/articles/2026-09-18/why-shoppers-are-turning-away-from-polyester-clothing?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS)
- ['Death Sentences': Crafting America's Favorite Countertops Is Killing Workers](https://www.bloomberg.com/news/features/2026-09-14/silicosis-lawsuits-mount-as-workers-seek-billions-in-damages?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS)
- [Trump's Quest to Be Consequential Puts the Entire World at Risk](https://www.bloomberg.com/news/articles/2026-09-18/iran-war-tariffs-and-trump-s-second-term-quest-to-be-consequential?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS)
- [Fender Is Making Enemies With a Messy Fight Over Its Iconic Strat](https://www.bloomberg.com/news/features/2026-09-17/fender-takes-on-the-electric-guitar-industry-over-stratocaster-copyright?utm_campaign=bw&utm_medium=distro&utm_source=yahooUS)

©2026 Bloomberg L.P.

[Terms](https://guce.yahoo.com/terms?locale=en-US)  and [Privacy Policy](https://guce.yahoo.com/privacy-policy?locale=en-US)

[Privacy Dashboard](https://guce.yahoo.com/privacy-dashboard?locale=en-US)

[More Info](https://finance.yahoo.com/more-info)

utiiiD

🟧 echo.blog ⭐

"Mirendil, an artificial intelligence startup launched by former Anthropic PBC researchers, is in talks to raise a new round of funding at a

Natasha Mascarenhas and Shirin Ghaffary (Bloomberg)

—

Interpretation history

2026-09-23T20:10:50Z
grounded: converges/medium — Converges: frontier-scale capital (Kleiner/a13, $5B) is now explicitly underwriting recursive self-improvement as the founding thesis — the same self-improving-
2026-09-23T13:28:23Z
case created — A Bloomberg-reported nine-figure round for an explicitly recursive-self-improvement lab is a consequential, resolvable financing episode with no matching open case.

Decision trace

09-24 06:10groundConverges: frontier-scale capital (Kleiner/a13, $5B) is now explicitly underwriting recursive self-improvement as the founding thesis — the same self-improving-loop architecture Scott has built framew
09-24 04:15createA Bloomberg-reported nine-figure round for an explicitly recursive-self-improvement lab is a consequential, resolvable financing episode with no matching open case.