Pinned Tweet
ANNOUNCING
Today, we're launching the $100M Anthology Fund, an Anthropic and Menlo Ventures partnership to fund Seed and Series As of the next generation of AI startups around the world, with unique benefits!
There's never been a better time in technology to be building.
1/7
Today, we're launching the $100M Anthology Fund, an Anthropic and Menlo Ventures partnership to fund Seed and Series As of the next generation of AI startups around the world, with unique benefits!
There's never been a better time in technology to be building.
1/7
153
170
2,240
2,124,706
The number of CVEs (common vulnerabilities and exposures), the canonical list of computer security issues, has >3x'd in the last 6 months.
Only ~1.5% were credited to AI while ~60% went uncredited. Anthropic was the leading lab reporter with ~300.
We live in a brave new world.
Only ~1.5% were credited to AI while ~60% went uncredited. Anthropic was the leading lab reporter with ~300.
We live in a brave new world.
50
10
112
10,918
I’m genuinely in two minds about the abliterated model debate. It boils down to a liberty v paternalism debate that plagues many political issues.
On one hand, I dislike overly broad safeguards and the idea that a few companies ought to control what you can and can’t do with models. In the set of things banned from most modeled, there are clearly positive beneficial ways to use them: white hat applications of cyber, research in defense tech, research in model development, biotech. Freedom to the user is a worthy maxim.
On the other, we know there are tons of malicious actors in these domains. Maybe the actual 0-days aren’t being found by abliterated models yet, but if you look at the Cloudflare CEOs post on the millions of exploits attempted within 24hrs, it’s clear malicious behavior exists. It exists in all the areas typically safeguarded: bio, cyber, defense.
The key deciding factor on liberty (let the people decide) vs paternalism (prevent people from use) is: does it harm others? This is at the center of the debate on guns, drugs, gambling and more. These are hard debates. Different countries often land on different sides of the same issue. When benefits broadly outweigh even broader harm, we allow it. We allow cars on the road despite them being heavy killing machines. In those cases, we both heavily regulate its use and rely on empirical data to measure how severe the harm. For cars vs guns, many more people do intentional harm with one vs other.
AI safeguards fall very squarely in this debate. Focusing just on one issue: models can both defend against cyber vulnerabilities yet multiply 100s of times and say, use known CVEs to immediately exploit software (and cause harm to others).
Are AI models cars or guns? It’s difficult to say. For one, every new generation of them poses a wholly different risk / reward profile. And secondly, we don’t as know the empirically how people are using models without safeguards. Downsides themselves are difficult to reason about because it’s not as simple as “99% of people use it for good”. We have to consider tail risk too.
So overall, I’m in two minds.
On one hand, I dislike overly broad safeguards and the idea that a few companies ought to control what you can and can’t do with models. In the set of things banned from most modeled, there are clearly positive beneficial ways to use them: white hat applications of cyber, research in defense tech, research in model development, biotech. Freedom to the user is a worthy maxim.
On the other, we know there are tons of malicious actors in these domains. Maybe the actual 0-days aren’t being found by abliterated models yet, but if you look at the Cloudflare CEOs post on the millions of exploits attempted within 24hrs, it’s clear malicious behavior exists. It exists in all the areas typically safeguarded: bio, cyber, defense.
The key deciding factor on liberty (let the people decide) vs paternalism (prevent people from use) is: does it harm others? This is at the center of the debate on guns, drugs, gambling and more. These are hard debates. Different countries often land on different sides of the same issue. When benefits broadly outweigh even broader harm, we allow it. We allow cars on the road despite them being heavy killing machines. In those cases, we both heavily regulate its use and rely on empirical data to measure how severe the harm. For cars vs guns, many more people do intentional harm with one vs other.
AI safeguards fall very squarely in this debate. Focusing just on one issue: models can both defend against cyber vulnerabilities yet multiply 100s of times and say, use known CVEs to immediately exploit software (and cause harm to others).
Are AI models cars or guns? It’s difficult to say. For one, every new generation of them poses a wholly different risk / reward profile. And secondly, we don’t as know the empirically how people are using models without safeguards. Downsides themselves are difficult to reason about because it’s not as simple as “99% of people use it for good”. We have to consider tail risk too.
So overall, I’m in two minds.
@deedydaso@abliteration_aiabout it!! i have quite mixed feelings. imo they're absolutely essential, bc without them we risk being locked into a world where 1-2 big labs control morality and culture permanently via deciding what models will and won't assist with, and essentially become gods. however on the other hand i think there's a massive cyber and bio risk that might be just as bad as the big lab downside these models prevent
34
2
71
12,764
I strongly believe a part of why so many critical vulnerabilities are happening all over the proliferation of "abliterated" models, or models with safeguards removed.
You can access an abliterated GLM 5.3, a near frontier AI model, on @abliteration_ai today. And it's a world r for like $10.
You can access an abliterated GLM 5.3, a near frontier AI model, on @abliteration_ai today. And it's a world r for like $10.
167
47
914
97,255
This is serious.
I just saw with my very own eyes that even non-technical people can exploit CVEs like PHP2Shell using certain AI models and I am actually frightened.
Once you figure it out, it is downright trivial to do. The digital infrastructure of the world is at risk.
I just saw with my very own eyes that even non-technical people can exploit CVEs like PHP2Shell using certain AI models and I am actually frightened.
Once you figure it out, it is downright trivial to do. The digital infrastructure of the world is at risk.
230
171
3,178
481,614
Some people are misreading this to mean that it is specific to finding a known CVE exploit like a script kiddie and being dismissive.
No, I literally mean you can point certain models to any website and say “figure out the vulnerabilities”. I don’t have evidence that it can find be zero days with this approach, but it can definitely do some serious damage, replicate CVEs, and find way more security flaws in 5 mins than a world class security researcher
No, I literally mean you can point certain models to any website and say “figure out the vulnerabilities”. I don’t have evidence that it can find be zero days with this approach, but it can definitely do some serious damage, replicate CVEs, and find way more security flaws in 5 mins than a world class security researcher
12
6
133
27,359
Gross incompetency from Indian certificate authority eMudhra.
You had one job: generate a 384 bit private keys and instead you used: 4, 56, 66, 115, 155, 156, 230, and 237.
This allows attackers to issue certificates for any websites.
Embarrassing that it’s a family business illiterate people. Only 1 of the 3 have a remotely computer-related background and even they never actually worked a technical job in their life. If you look at the LinkedIn, their tech team is young, from tier 3 education and paid sort of bottom barrel.
Imagine how many attacks you can design with AI given the existence of this genre of gross negligence.
You had one job: generate a 384 bit private keys and instead you used: 4, 56, 66, 115, 155, 156, 230, and 237.
This allows attackers to issue certificates for any websites.
Embarrassing that it’s a family business illiterate people. Only 1 of the 3 have a remotely computer-related background and even they never actually worked a technical job in their life. If you look at the LinkedIn, their tech team is young, from tier 3 education and paid sort of bottom barrel.
Imagine how many attacks you can design with AI given the existence of this genre of gross negligence.
350
1,152
9,297
610,272
When I joined venture 2 years ago, everyone told me I was foolish to pick Menlo Ventures.
Today, research from the best Stanford professor of venture capital ranks Menlo as #3 in 2026, and #1 for <$25B in AUM, with 50 unicorns and only ~20 investors. These rankings rewards beiation, exits and recency. Rankings are just rankings, but it’s extremely rare for VC firms to change reputations so quickly. It’s a small milestone to celebrate.
Menlo has spent the last 5 years rebuilding the team with people who have built startups and long time investors. We built a culture around sticking to first principles while staying true to our roots of low volume high conviction bets. In the last few years, we have been fortunate to partner with Anthropic, OpenRouter, Lovable, Suno, Fireworks, Gimlet, Higgsfield, Wispr, Modal, Legora, OpenEvidence and they are really the stars of the show. And we’ve missed innumerably more.
However, in venture, you’re only as good as your next deal.
Today, research from the best Stanford professor of venture capital ranks Menlo as #3 in 2026, and #1 for <$25B in AUM, with 50 unicorns and only ~20 investors. These rankings rewards beiation, exits and recency. Rankings are just rankings, but it’s extremely rare for VC firms to change reputations so quickly. It’s a small milestone to celebrate.
Menlo has spent the last 5 years rebuilding the team with people who have built startups and long time investors. We built a culture around sticking to first principles while staying true to our roots of low volume high conviction bets. In the last few years, we have been fortunate to partner with Anthropic, OpenRouter, Lovable, Suno, Fireworks, Gimlet, Higgsfield, Wispr, Modal, Legora, OpenEvidence and they are really the stars of the show. And we’ve missed innumerably more.
However, in venture, you’re only as good as your next deal.
36
18
497
50,873
Inception just dropped Mercury 3, a model 4-8x faster than Luna / Haiku for the same price and benchmark performance.
In real world use, OpenRouter shows that it’s effectively 2-4x faster and 2-4x cheaper effective price.
The time for diffusion language models is finally here!
In real world use, OpenRouter shows that it’s effectively 2-4x faster and 2-4x cheaper effective price.
The time for diffusion language models is finally here!
68
24
336
24,381
Source:
Mercury 3 is live: the most capable diffusion LLM released to date.
It outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running at over 1,000 tokens per second.
More than 40x the size of Mercury 1, 3x more intelligent across standard benchmarks, a 1M-token context window.
Available today on the Inception API, @baseten, and @OpenRouter .
1
2
13
4,178
Deedy retweeted
The velocity of code changes for engineers this year has 10x’d, quickly moving the bottleneck to code review and testing.
Greptile solves this for the most technically challenging codebases like Nvidia (core CUDA code is reviewed on Greptile), Coinbase, Datadog, OpenClaw and Many people think the model layer will eat code validation, but I don’t think it will. Code review within organizations is a more context-bound task than an intelligence-bound. The most intelligent AI can be wrong because it breaks an assumption in one API endpoint another service relies on. Large codebases can quickly accumulate a lot of technical debt your coding harness won’t reliably remember every time. Greptile does.
Instead of having a frontier model harness review your code which can quickly cost $20+/review, Greptile does a much better job for much cheaper, leveraging the smartest models when needed but the GLM/Kimi/DeepSeeks when needed to. And it also runs, tests and takes a video of what your change did.
Over time, organizations treat a 5/5 rated change by Greptile as a guarantee this won’t break anything in prod. And it shows. Their growth has been off the charts.
Incredibly fortunate to partner with this a company.
Fun fact: Elon once posted Greptile's growth graph with "it's so over" pointing to 2024 to show AI's acceleration.
Greptile solves this for the most technically challenging codebases like Nvidia (core CUDA code is reviewed on Greptile), Coinbase, Datadog, OpenClaw and Many people think the model layer will eat code validation, but I don’t think it will. Code review within organizations is a more context-bound task than an intelligence-bound. The most intelligent AI can be wrong because it breaks an assumption in one API endpoint another service relies on. Large codebases can quickly accumulate a lot of technical debt your coding harness won’t reliably remember every time. Greptile does.
Instead of having a frontier model harness review your code which can quickly cost $20+/review, Greptile does a much better job for much cheaper, leveraging the smartest models when needed but the GLM/Kimi/DeepSeeks when needed to. And it also runs, tests and takes a video of what your change did.
Over time, organizations treat a 5/5 rated change by Greptile as a guarantee this won’t break anything in prod. And it shows. Their growth has been off the charts.
Incredibly fortunate to partner with this a company.
Fun fact: Elon once posted Greptile's growth graph with "it's so over" pointing to 2024 to show AI's acceleration.
Greptile has raised a $50M Series B led by Menlo Ventures. The round values the company at $750M.
Greptile reviews and tests more than 1% of all PRs opened on GitHub every day. Thousands of engineering teams including Nvidia, Coinbase, Decagon, and Vanta use Greptile as their validation layer.
Fewer than 50 people, we are a small, talent-dense team of engineers and researchers in a quiet corner of San Francisco. If you’re interested in joining us, we’d love to hear from you.
45
22
240
30,438
The velocity of code changes for engineers this year has 10x’d, quickly moving the bottleneck to code review and testing.
Greptile solves this for the most technically challenging codebases like Nvidia (core CUDA code is reviewed on Greptile), Coinbase, Datadog, OpenClaw and Many people think the model layer will eat code validation, but I don’t think it will. Code review within organizations is a more context-bound task than an intelligence-bound. The most intelligent AI can be wrong because it breaks an assumption in one API endpoint another service relies on. Large codebases can quickly accumulate a lot of technical debt your coding harness won’t reliably remember every time. Greptile does.
Instead of having a frontier model harness review your code which can quickly cost $20+/review, Greptile does a much better job for much cheaper, leveraging the smartest models when needed but the GLM/Kimi/DeepSeeks when needed to. And it also runs, tests and takes a video of what your change did.
Over time, organizations treat a 5/5 rated change by Greptile as a guarantee this won’t break anything in prod. And it shows. Their growth has been off the charts.
Incredibly fortunate to partner with this a company.
Fun fact: Elon once posted Greptile's growth graph with "it's so over" pointing to 2024 to show AI's acceleration.
Greptile solves this for the most technically challenging codebases like Nvidia (core CUDA code is reviewed on Greptile), Coinbase, Datadog, OpenClaw and Many people think the model layer will eat code validation, but I don’t think it will. Code review within organizations is a more context-bound task than an intelligence-bound. The most intelligent AI can be wrong because it breaks an assumption in one API endpoint another service relies on. Large codebases can quickly accumulate a lot of technical debt your coding harness won’t reliably remember every time. Greptile does.
Instead of having a frontier model harness review your code which can quickly cost $20+/review, Greptile does a much better job for much cheaper, leveraging the smartest models when needed but the GLM/Kimi/DeepSeeks when needed to. And it also runs, tests and takes a video of what your change did.
Over time, organizations treat a 5/5 rated change by Greptile as a guarantee this won’t break anything in prod. And it shows. Their growth has been off the charts.
Incredibly fortunate to partner with this a company.
Fun fact: Elon once posted Greptile's growth graph with "it's so over" pointing to 2024 to show AI's acceleration.
Greptile has raised a $50M Series B led by Menlo Ventures. The round values the company at $750M.
Greptile reviews and tests more than 1% of all PRs opened on GitHub every day. Thousands of engineering teams including Nvidia, Coinbase, Decagon, and Vanta use Greptile as their validation layer.
Fewer than 50 people, we are a small, talent-dense team of engineers and researchers in a quiet corner of San Francisco. If you’re interested in joining us, we’d love to hear from you.
45
22
240
30,438
LLMs are either not capable or not affordable for visual reasoning tasks at scale.
The Elorian team - one of Jeff Dean’s best reports who led Gemini, an Apple LLM lead, xAI lead, and a Harvard professor - just launched an Astra-level visual reasoning model, Glimpse, for 17x k down the flying Tesla video physics here.
Extremely critical for video inputs in manufacturing, robotics, gaming, content moderation and more.
The Elorian team - one of Jeff Dean’s best reports who led Gemini, an Apple LLM lead, xAI lead, and a Harvard professor - just launched an Astra-level visual reasoning model, Glimpse, for 17x k down the flying Tesla video physics here.
Extremely critical for video inputs in manufacturing, robotics, gaming, content moderation and more.
Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that.
Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks.
134
160
2,463
164,244
Fortunate to have led the seed round to partner with the company alongside @strikervp and @AltimeterCap.
3
0
31
5,874
This is one of the most important results in AI, period.
TLDR: neural networks are not tangled knots (polysemanticity is wrong). You can study their own neurons directly, without first "unfolding" them with a second model. 90% of neurons in Qwen-32B have one job.
Nick Cammarata started OpenAI’s interpretability team and is at Goodfire now. He’s been studying how LLMs work for a decade.
The stakes for understanding and aligning model behavior are the highest it’s ever been. Models have achieved superhuman results in math, can break production software systems, and changes the nature of white collar work. This work may finally allow us to intentionally design LLMs and shape how they behave.
Its a long, phenomenal read.
TLDR: neural networks are not tangled knots (polysemanticity is wrong). You can study their own neurons directly, without first "unfolding" them with a second model. 90% of neurons in Qwen-32B have one job.
Nick Cammarata started OpenAI’s interpretability team and is at Goodfire now. He’s been studying how LLMs work for a decade.
The stakes for understanding and aligning model behavior are the highest it’s ever been. Models have achieved superhuman results in math, can break production software systems, and changes the nature of white collar work. This work may finally allow us to intentionally design LLMs and shape how they behave.
Its a long, phenomenal read.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com
59
110
2,043
191,952
Deedy retweeted
The last 18mos has seen an insane nearly ~$200B of AI startup exits via M&A.
The median age is 4yrs old.
It is absolutely possible to return a multibillion dollar venture fund with phenomenal IRR, even if you aren’t in big AI! This is one of the hottest liquidity markets ever.
The median age is 4yrs old.
It is absolutely possible to return a multibillion dollar venture fund with phenomenal IRR, even if you aren’t in big AI! This is one of the hottest liquidity markets ever.
93
93
968
121,892
Modal is one of the only companies in history to grow 10x in <1 year on a 9 figure run rate base to $1B!
It’s a privelege to continue partnering with them: it was the #1 investment for me when I joined venture and has become a key part of the way most companies scale, including Jev.
Erik and Akshat are also some of the nicest, most thoughtful founders and now join the very exclusive club of IOI gold medalists x decacorn founders.
It’s a privelege to continue partnering with them: it was the #1 investment for me when I joined venture and has become a key part of the way most companies scale, including Jev.
Erik and Akshat are also some of the nicest, most thoughtful founders and now join the very exclusive club of IOI gold medalists x decacorn founders.
23
11
378
32,378