ex-@GoogleDeepMind. Used to work on Gemini. Taking a sabbatical to explore my curiosity.
New York, USA
Joined March 2013
- Tweets3.2K
- Following841
- Followers1.3K
- Likes29.5K
This will definitely only be used to get sick grammable shots while trail running, and nothing else.
Purdue University's IDEAS Lab built an autonomous drone that tracks runners through dense forests completely without GPS or cloud networks.
By running 3D LiDAR mapping and obstacle-avoidance AI entirely on-chip, it can weave under tree canopies and predict target movement through blind spots in real time.
Massive implications for next-gen search & rescue and edge AI.
Varun Godbole retweeted
Tbh I’d love to see more AI products help people develop discernment about what’s worth outsourcing, rather than treating every new use case as progress. Product and design decisions shape what people come to see as natural uses of a technology
Varun Godbole retweeted
the tech industry spent basically the past 15 years doing the worst possible thing for their users and now LLMs are letting people debloat their TVs, ignore upsells in checkouts, revive games studios gave up on, crack open whatever software you want
i love it
Varun Godbole retweeted
The beginning of Spellbook was also an oddly meditative and spiritual pursuit: Our hypothesis was that if we could access raw reality directly, and disconnect from the sort of "false reality" of popular stories, then we would find alpha in the world.
In some ways the company was a higher level experiment: if we can come to see raw reality with crystal clarity, then alpha should follow. Alpha exists in the gap between popular stories, and raw reality.
We saw business as competitive philosophy. Your score is a direct measure of how good you are at uncovering truth before others.
This continues to provide a great deal of meaning for me: investing and business are a gym that ultimately enable you to build radically clear thinking.
It worked: we raised money for a legaltech company long before it was fashionable (pre-Atrium), and were the first in the world to commercially launch and LLM-based tool for lawyers.
We were making a hard bet that legal AI was needed, based on raw first principles thinking, long before any story had taken hold.
The challenge, though, is that the "raw reality" view that uncovers alpha, is a completely different from the story-telling muscle that excites customers and your team.
The hardest part of business is that you must simultaneously be aware of stories as a pollutant, while still crafting the inspiring stories that ultimately drive most people to action.
Investors/hedge funds have the luxury of only needing to be good at seeing raw reality. They don't need to market their bets. Sales and marketing has the luxury of being able to focus on the stories.
There are very few people who can simultaneously wield the power of stories, while also piercing through the sea of stories, extracting alpha from the raw reality that eludes others.
The ultimate magic is being able to do both.
Realizing how incredibly autistic our early board meetings were for Spellbook:
1. I declared the meeting a “story free zone” because I was worried that stories were too psychologically persuasive, and that they would pull us away from hard data-driven reality
2. Meetings would largely consist of scrolling through like 40 slides of charts sifting for signal. I called this “soaking in the data”
3. In the search for PMF I wanted us to think like a hedge fund making many bets. While we knew the problem we were solving I didn’t want our team to get too tied to any single narrative because I thought it would reduce agility.
4. I never allowed us to think about us as “a company”, but instead as a SWAT team searching for a company to build
5. This approach enabled us to launch over 100 different landing pages with different angles on the way to PMF
Honestly, I think it’s a really good approach pre-PMF. Too many companies get attached to a singular story too early and quit when it doesn’t work out. Startups are 3D chess. There is always a move. Slack was a game company.
However, at Series B/Series C scale you really need stories to align both your team and customers, and you can think much more longterm after PMF.
But I still believe stories are a dangerous dark magic. You can craft an incredibly persuasive story for almost any position. They’re more of a marketing/comms tool than a reasoning/understanding reality tool.
Varun Godbole retweeted
so when everyone thought Facebook was spying through your phone it turns out tv makers were the ones doing it all along, this is a nightmare
Varun Godbole retweeted
I am also optimistic that AI/cyber will be defense-friendly after things stabilize, but before that it will be offense-friendly, which creates a lot of dynamics that are difficult to model:
1. For any existing or new vulnerabilities discovered by governments and other offensive actors, there will be an incentive to exploit them quickly, rather than sit on them quietly, because they will anticipate that the defender will find the vulnerability and patch it much faster than before.
2. The surface area of cyber is very broad, and finding vulnerabilities is different from fixing them, which is also different from fixing them without also introducing new, even subtler security problems into the code.
3. We do not know how high the ceiling for subtle bugs might be. Do there exist vulnerabilities that were simply too hard for humans to find or laborious to exploit? How many vulnerabilities? How long will they persist? We are now getting to the point where we have to reason about systems with quasi-superhuman cognitive performance. This is hard! Counter-intuitive things will happen.
4. There are some threats such as hardware vulnerabilities or new discoveries in cryptography where the offense/defense dynamics could be importantly different from traditional cyber.
There was always going to be some tumult as models gained cyber capabilities. This was by far the easiest major novel AI risk to project, in broad terms. It may be true that broad and rapid diffusion is the optimal path forward; that outcome wouldn’t surprise me that much. But it could also be true that, given the offense-favoring dynamics at first and the uncertainties I have described, it’ll be optimal to slow diffusion. And regardless of whether it is optimal to do so, I would expect substantial political pressure to “do something” if the cyber events are sufficiently damaging or disruptive.
The optimal outcome is one where we diffuse fast and comprehensively enough to find and fix vulnerabilities while also not causing too much tumult. Maybe we are striking the balance effectively now, as a global industry! I hope so. But I also hope everyone is attentive and ready to update their beliefs/strategies as we gather empirical data. I’m skeptical of anyone who seems confident here in either direction.
Near as I can tell, despite a lot of scary potential, nothing truly catastrophic has happened yet (the water is running; the lights are on, no cataclysmic data breaches I am aware of). I hope that continues to be the case. I do agree with the AI risk skeptics that there is a lot of catastrophizing on X about this, and the current attacks, while significant, seem well within the range of things we can bounce back from.
If things continue at this level of severity, I feel pretty good about our situation with AI/cyber. But if it gets noticeably worse and seems to be on a trend of getting even more severe, which is my expectation, we could be in for trouble. Maybe society can digest way more cyberattacks than I had apprehended. Maybe “cybersecurity is not very important” as the famous saying/paper goes. I’m very much open to being pleasantly surprised.
yeah as others have said, two different trends around AI + cyber are happening at the same time and its confusing:
AI helping cybersecurity situation:
- 5.6 sol (yay daybreak) discovering this PHP vulnerability that was already there so people can patch it (of course can be worse in short term while lots of systems go unpatched, and also long tail of legacy unpatched systems)
AI not helping cybersecurity situation:
- lots of AI powered automated cyberattacks (e.g. railway and opencode hacks)
Vulnerability discovery probably on average favors defenders in the medium term. Automating cyber operations will help attackers a lot though, and probably enough to overwhelm positive effects on the former (every random community bank or old person is not gonna have AI swarms protecting them against AI swarm attacks). Of course the cycle of patching vulnerabilities is becoming a lot harder because of how fast automated cyberattacks are.
Also an aside on AI spear fishing in particular -
Literally today I had one of those fake tech crunch reporters reach out over DMs and the fake was pretty convincing (thousands of followers, tweets and photos about tech crunch, but when I googled became clear person didn't exist).
People who are not used to the idea that people can lie in convincing ways on the internet are going to get exploited over and over again and its gonna be bad (and also generally make it way harder to have good interactions with humans you don't know online).
Varun Godbole retweeted
This is truly getting tiresome. The effort to paint the very difficult immediate situation as without precedent and not “normal” ignores the waves of the security issues of the past.
I’m not downplaying this and just as tired of “no, this time is different.”
There’s just no escaping that every step function change in automated tools to find (security) holes/bugs that arose out of point experiences being generalized immediately created the type of “all hands” situation we face now.
Everything from floppy disk scanning to buffer over runs to uninitialized variables to off by one loops to automation hooks to script kiddies to xss to frames to social engineering saw incidents leading to generalized tools that scanned environments to find countless latent bugs that needed to be fixed urgently.
One day a product was ready to ship and then a tool ran and suddenly there were 10,000 new bugs for 500 engineers to investigate and fix and then test for what broke because well who knows what happens when you suddenly initialize a variable that wasn’t previously.
This sucks in a ton of ways. No one wants to stop what they are doing at what seems like the biggest opportunity and/or disruptive risk to business in a generation. No one wants to touch old code. No one likes to have work generated for them by automated tools finding issues that are obscure by definition.
But this is all good. These bugs were all there. Instead of cursing just realize how lucky your team is to find them now and not after the breach. Whenever we surfaced from these moments in the past we felt our code reached a new and better “steady state” and that will happen.
Varun Godbole retweeted
Such a great post! Insights I particularly appreciated:
- "Why hasn’t a huge discontinuity in damages happened since the dawn of scaled AI? This non-responsiveness of damages to a rising tide of AI capabilities is the major conceptual crux, and stumbling block, of all AI security policy today."
- "cybercriminals are unskilled and get arrested a lot ... Criminal organizations need to base themselves in middle and lower income countries, where they can pay less and hire people who have few formal-economy options. ... I think folks in the AI safety and cyber world sometimes imagine adversaries as their Waluigi twins, implicit MIT graduates; in fact, in cybercrime we’re dealing with a lower-skill demographic."
- "Criminals also contend with the risk and reality of arrest, sanction, interdiction, and prosecution, which caps their ability to scale: the larger the organization, the more domestic and international law enforcement attention it attracts."
- "It took the attacker ecosystem a very long time to mature the workflows, business models, and industrial divisions of labor that scaled ransomware into a multi-billion dollar industry, and it’ll take a long time to develop new models that centrally leverage AI throughout criminal labor processes... it took 5-8 years for ransomware to become a huge deal"
- "As risk professionals we should assume this will all change and that what we have now is a defender-dominant window in which we can prepare for some very bad possible futures"
Lots of interesting charts too: joshuasaxe181906.substack.co…
Contrary to the implicit mental models of many on this website, cybercriminals aren't the security-engineer Waluigis of Stanford-educated big tech engineers, or the criminal analogs of a16z-backed AI pentesting startups.
They're generally less talented, recruited from middle-income countries with weak legal oversight and fewer legal tech job options, and struggle with arrests, sanctions, retaining talent, and stable access to AI inference. Ransomware took years to become a scaled business under these conditions, and AI-native organized criminal syndicates that fully leverage the power of AI to scale may take awhile too.
Nation-states, meanwhile, are quick to adopt AI for vulnerability research and exploit development, but historically slow to deeply retool military operations with new technologies. And potentially devastating agentic cyber swarms may become weapons in their arsenals alongside e.g. ballistic missiles that they possess but don't immediately use.
So to those calling php2shell "the worst day in cyber ever" and expecting immediate spikes in cyber damages: there isn't the direct relationship between new capabilities and new damages you're imagining.
This is good news, obviously. It means defenders have time to prepare (and we should be shouting from the rooftops that we're in a defense-advantaged interregnum, currently). Huge risks are approaching but at the pace of technology diffusion into less innovative parts of the world economy and state sector, not at the pace of AI capability growth.
Full version: joshuasaxe181906.substack.co…
Varun Godbole retweeted
no but every instance of this so far has been fixable by disabling some hardware feature (like down clocking or disabling hyperthreading)
Is it just me or is it a terrible idea to build something that can convincingly pass off as a human to create affordances for emotional attachment? This seems like AI psychosis fuel?
Meet Dessa, our speech model trained from scratch on human conversation.
Hear it for yourself: block.xyz/dessa
Varun Godbole retweeted
So that's terrifying but probably true.
What do you do, then, stop using Node, Python, and PHP completely? Vibe-code your own frameworks and go with security through obscurity?
It's getting pretty alarming that I have to spend an hour a day hardening my security and I'm just some joe with a couple of agents. I can't imagine how exposed companies are right now.
Varun Godbole retweeted
"It is in fact possible to take an entire category of critical infrastructure off the internet entirely. With a fully self-contained network where all the endpoints are off grid. This is not a new concept-- anyone could do this for your home today with equipment you can buy from Best Buy."
The cyber incidents of the last 48 hours have been pretty mind blowing. I’ve been trying to piece together the “why now” question, and I think what we’re seeing is the first of a new kind of super-agent driven vulnerability cascade-- first one vulnerability is found by a malicious actor, then other bad actors pop up to try and take advantage of it before companies patch their systems.
But because each bad actor is using swarms of thousands of super-agents, those agents are not only exploiting the existing vulnerability, but finding new zero days as well, all in a matter of hours. And the cycle repeats. As of the writing of this post there have been vulnerabilities found across at least half-dozen independent systems including LiteLLM, PHP, eMudhra Certificates, NGINX, Bind 9, and NodeJS. The list continues to grow hour by hour.
Basically everything connected to the internet is at risk. Not just company websites and the usual databases of sensitive personal information, but all our critical infrastructure-- banks, power grids, pipelines, government agencies, all of it.
The good news is that cybersecurity and ops teams had a chance to see this coming. This is why from when Mythos first dropped and had our first warning, these teams have been hard at work shoring up systems across the board.
This is also why at Astranis we’ve seen a massive uptick in demand for truly air-gapped, sovereign networks, deployed via dedicated satellite to span an entire country or connect a large enterprise.
It is in fact possible to take an entire category of critical infrastructure off the internet entirely. With a fully self-contained network where all the endpoints are off grid. This is not a new concept-- anyone could do this for your home today with equipment you can buy from Best Buy. But to stand up such a network covering a large geographic area or an entire country, without touching the existing fiber or other network infrastructure, you can only do that with a dedicated satellite. And specifically a satellite in a special orbit called geosynchronous orbit.
Over the past year we’ve begun working with some of our customers on what it will take to implement truly air-gapped, standalone networks in support of mission critical use cases like core government functions, emergency response, defense, core infrastructure, or all of the above.
That might sound extreme, but that is the kind of measure we expect will be required to shrink the attack surface of some of the world’s most critical infra. That’s just the position we are now in. It's a brave new world.
I think a lot of secrets of various sorts are about to be rapidly and potentially painfully unveiled to the world. The security landscape in the next few years is going to be wild.
so for all the people that think their phone has been listening and targeting ads even though it is almost impossible, it just might have been the tv in the room and that is basically verified xitter.cf/0x0SojalSec/status/211…
Agents be agenting
‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’
Varun Godbole retweeted
Alcoholic cybersec friend at Fortune 500 insurance company texted to tell me he received an email from his boss thanking him on his diligence patching every PHP box overnight but he was drunk all day yesterday and didn’t know anything until waking up 30 minutes ago
