What’s 🔥 in Enterprise IT/VC #469
Karpathy’s Agent Reality Check and Why It’s Great News for Builders + highlights from the Inaugural AI Security Summit
Last week’s Andrej Karpathy interview on Dwarkesh got everyone’s attention as he said AGI is still a decade away. More specifically on AI agents, Karpathy said “they just don’t work” and “don’t have enough intelligence, they’re not multimodal … they don’t do computer use … they don’t have continual learning.”
On the one hand, this can be incredibly bearish, but on the other hand, when you dig through what Karpathy is really saying, agentic workflows still can be beautiful, especially when paired with humans versus the promise of full autonomy. While current “agent” systems (as conceived) still have major gaps and the time-horizon for full autonomy is at least a decade away, there are still massive opportunities in human + agent collaborative workflows. I, for one, view the glass half full and believe we’re just early in the autonomous future, and there is so much to be built to get us ready as I highlighted in our boldstart fund vii announcement focused on the “autonomous enterprise” earlier this summer.
The good news is that we have companies like Walmart and BNY continuing to plow ahead and sharing that the tech actually does add value.
As I see it, there is so much that startups and existing vendors can build to get to this autonomous world which include many highlighted in the Karpathy interview:
Memory & Context → Karpathy said “you can’t just tell them something and they’ll remember it”, pointing to the lack of continual learning and persistent memory as a major blocker.
Tool Use & Interfaces → Noted that today’s agents “don’t do computer use” and can’t reliably interact with real systems or UIs.
Reasoning & Planning → Called current agents “cognitively lacking,” with errors compounding over multi-step tasks.
Orchestration & Reliability → Described how chaining actions makes failure rates explode, pointing to the need for better infra, control loops, and collaboration with humans.
Security & Identity → repeatedly stressed that agents must be trusted, auditable collaborators, not unpredictable black boxes. (This aligns directly with why infra like identity & permissions will be essential.)
So the timing couldn’t have been any better as Snyk and AI Engineering (Swyx) hosted the first AI Security Summit in San Francisco this past week with leadership and practitioner tracks. The energy was electric, and it was so awesome to see this community of vendors and some of the largest customers like JP Morgan and RBC share the importance of security from inception and design, especially when it comes to agents. Here’s a shot from a panel I was on with fellow VCs Guru Chalal from Lightspeed, Chenxi Wang from Rain Capital, and Taher Elgamal from Evolution Equity Partners.
If interested, here’s an interview I did on the show floor and corresponding story in Silicon Angle about the many opportunties to secure agents and AI.
Most panels and discussions centered around guardrails, governance and agents and authorization. The launch of Snyk’s Evo was definitely a highlight of the conference as it launched the first swarm of security agents to automate threat modeling, red teaming, and remediation to secure AI-native applications and can automatically create fixes, JIRA tickets, and Pull Requests.
I’m biased, of course, as an investor in Snyk, but also one other company that generated a ton of buzz at the event, Keycard which just emerged out of stealth. I’m super fired 🔥 up to have led the Inception round with my friends at a16z and double down in the A with Acrew Capital who we also did Protect AI with.
And here’s why it’s so important for the agentic era:
To summarize while we’re more than a decade away to fully autonomous agents, a world that I assumed would at least take that long, the opportunity to build and deploy agentic workflows is still incredibly massive as we build towards that future!
As always, 🙏🏼 for reading and please share with your friends and colleagues!
Scaling Startups
#Brian continues to share his thoughts about potentially being at the top…
#💯 and ties in nicely with last week’s What’s 🔥
What’s 🔥 in Enterprise IT/VC #468
The Many Paths to $100M ARR: Understanding PLG, SLG, and BLG Tradeoffs (Plus: Varun from Clay's 3 lessons for AI founders building for enterprise)
#the importance of ball control and being able to lead and set terms for funds…
#
Enterprise Tech
#a must read from Alex Konrad, Upstarts - the deep kind of reporting we need in this era, click below to read story directly and sign up!
#👀 still need labelers and humans but this is pretty intense - junior banker rolls will disappear quickly
#not to be outdone, Goldman Sachs has already built their own model
#more dystopian thoughts 🤖
#yes, this is what I’m watching in the boldstart portfolio as well
#Anthropic continues to 📈 and 80% of revenue from enterprises!
#major breakthrough from DeepSeek…again - turning text into vision tokens using fraction of tokens an LLM needs to realize 97% accuracy
#🤯 “Replit credits much of this growth to higher-margin business customers like Duolingo and Zillow.”
#congrats Baz, a portfolio co, and great list of top AI startups!
#agent platforms on 🔥 as Langchain raises $125M at $1.25B on heels of n8n raising at $2.5B
#how the next AWS outage can be prevented
#for those interested, so many more enterprise and real-world use cases like tokenization of assets coming from oldest and largest financial institutions
Markets
#great slides from Coatue - are we in a bubble or not?
#speaking of agents and security, this is a huge price for Securiti AI - notice “help customers ensure that AI agents and models don’t access data that should be private”
#Not surprising…
