Ed Sim (@edsim) header
Ed Sim (@edsim) profile picture

@boldstartvc: the Inception fund. Helping the bold start. First check for technical founders. Security, AI infra, physical AI, agents & the weird. What’s 🔥 👇

Miami
Joined January 2009
Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)Photo by Ed Sim (@edsim)
🔥 up to announce @boldstartvc Fund VII $250M to back bold technical founders building the autonomous enterprise. From Inception. Before the world believes. It always starts with an idea that feels insane… until it isn’t. 🎥👇🧵
106
38
22
485
94,560
just imagine the security nightmare 🤯 AI rewriting what someone says live and matching their lips in real time… only a patent filing for now, but deepfake fraud just got a lot easier to imagine
Microsoft has filed a patent application for AI that can rewrite what people say during livestreams in near real time and change their lip movements to match. The system would flag words based on preset rules or a viewer’s profile, such as age, language or location. Generative AI would then replace or remove those words, smooth the audio and change the speaker’s mouth and facial movements so the edit is seamless. Examples in the filing include live translation, cutting filler words and stutters, fixing mispronunciations, removing profanity and adapting content for different age groups or regions. The filing says the original unedited stream would be kept as a “ground truth” record. Microsoft hasn’t said whether it plans to build the product.
1
9
756
hats off to Anthropic and SpaceX, legendary 🚀 venture has changed forever... Not only the outcomes but also the capital and speed $9B raised used to be headscratching but that was done in 24 years versus $100B in under 6 🤯 Question is what happens when everyone underwrites the next one to look just like them, a ton of great businesses worth less than these immortal opps will get missed! What's 🔥 #525 whatshot.vc/p/whats-in-aiinf…
1
2
12
1,248
100x is coming faster than most think 👀 all of our infra was built for humans, now it has to be rebuilt for 100x agents, starting with access
"In February, machines outnumbered humans in terms of token consumption. And then I think by September it was 7.8 times. And so, how long will it take before we hit 100 times, right? You have to support that scale, then you've gotta do dynamic access at that scale" - @edsim Founder and General Partner @Boldstartvc
2
3
15
1,162
A fun one and we covered a lot of ground in 15 minutes - listen at the end for some startup building anti patterns Thanks for having me @insecureagents @vtahowe at @AISecSummit
Access Built for Humans Won't Survive Agent Scale, with Ed Sim (Boldstart Ventures) "You've gotta do dynamic access at that scale, a millisecond scale, with a thing that's growing 10x, 20x year over year." Ed Sim (@edsim), Founder & General Partner at @Boldstartvc, joined us live at @AISecSummit. Agents passed humans in token consumption in February, and by September they were at roughly 7.8x. Ed's argument: access granted once and held forever was built for humans, and every layer of the stack has to be rebuilt for agents. We get into: > Why buyers want one central policy across every vendor's agent > Scoping an agent to one database, for one job > Investing two years ahead: sovereign cyber models, AI data centers, physical AI > Founder anti-patterns when speed is the only moat TIMESTAMPS (00:00) A year after the last AI Security Summit (00:40) Personal agents as the test bed for agent security (01:40) Agents are relentless: 200 API calls per hour against a reservation site (02:50) Every agent gets its own identity, and why banks balked two years ago (03:50) The fishbowl problem: one identity across every vendor (04:50) One place for a central policy (05:50) One database, one job, no forever access (06:10) Boldstart Fund Seven and the autonomous enterprise (06:45) Agent tokens pass humans: 7.8x by September (07:40) OAuth islands and long-lived credentials (08:30) Investing two years ahead, and why large $100M inception rounds don't always gaurentee teams win (10:30) Sovereign cyber models, AI data center security, and physical AI (12:40) Misfit toys: betting on the non-obvious (13:50) Founder anti-patterns: sell earlier, speed is the moat, go multi-product
2
1
11
949
Automated fixing is only as good as what you feed it. @snyksec is one of the first partners in OpenAI's Daybreak Defense Factory, sending real, prioritized findings into the Discovery stage. Fixing a false positive fast is still wasted work. teapawt.org/snyksec/status/2123060…
OpenAI just announced Daybreak Defense Factory, an automated defense operation that finds, validates, and fixes vulnerabilities continuously. Snyk is among the first integration partners. Our findings enter at the Discovery stage, and that placement matters. Automated fixing is only as good as the findings feeding it. A fast fix for a false positive is wasted work. A fast fix for the wrong problem is worse. An agent needs to know what’s real and what to fix first. Snyk brings more than a decade of security context to that question, and findings can arrive with dependency paths and remediation details. That’s what we’re putting into the Defense Factory. See how to set up the integration: vist.ly/5tc4g
1
2
346
NO!!! This would be incredibly damaging to our startup ecosystem and limit choice Assuming we go this path who is going to offer the best sovereign open weight models? Very few can afford to build and offer.
SITUATION DETECTED: The White House and Treasury are discussing a proposal that would require US companies to publicly disclose the use of Chinese open-weight AI models on their websites, aiming to discourage their use through public pressure, per Reuters.
4
2
18
1,907
My greatest fears getting closer to being realized
Nvidia fallout... Will all roads lead to the largest labs? My fear on a potential Open Weight Squeeze A risk, not a prediction. What’s 🔥 👇 #524
2
215
Hard agree but would be killer if you made them open weight! So much opportunity for whoever has the best sovereign open weight model
We’ve heard again and again that enterprises don’t want to pour all their valuable IP, data and tacit knowledge into someone else’s model. They want control of their own weights.
6
1,030
Sure seems that way and glad OpenAI is leading the way Amazing at finding vulnerabilities but still needs context and deep understanding of systems to patch at scale. Also seeing lots of complaints on cost and still handicapped so can’t defend the way one needs to… Which means so much opportunity for swarm and intelligence first startups. Market is massive, problems are serious, not winner take all, and will be layered with different players. Keep building 🏗️. Lots to do to keep us all safe 🙏🏼
2
4
1
8
1,530
Proteins move, and the best drug discovery models need to learn how. @ToposBio is now running 2-3x more molecular dynamics with @nvidia cuPotential, and up to 7x with its own optimizations. More physics means better training data for its generative models, which is how you go after targets once considered undruggable 🧬 Proud to back @r_zarcone and @khosra from inception. teapawt.org/ToposBio/status/212243…
Proteins move. Drug discovery moves with them. With @nvidia and @NVIDIAHealth cuPotential in our OpenMM workflows, we're running ~2-3× more molecular dynamics. Up to ~7× with our own optimizations: more physics to train our generative models on. More detail on our blog: toposbio.ai/blog/nvidia-cupo…
1
1
300
Sadly yes. We can’t let this happen Wrote more about it my weekly newsletter Saturday What’s 🔥 #524 👇🏼
The Financial Times acknowledges that the ouster of Nvidia CEO Jensen Huang is a massive blow to the open-source community. This is a victory for the aspiring monopolists at Anthropic and OpenAI, which have been lobbying the US government to restrict Chinese open-source models -- exploiting "national security" rhetoric to try to ban their main competitors. Link: ft.com/content/418ca885-fbcf…
3
2
1
7
937
Yep. Been a long time coming. There is no time to wait. The time is now. You have to fight compute with compute. Cybersecurity has turned upside down and so much opportunity for swarm and intelligent first new companies teapawt.org/edsim/status/208582992…
Hackers are now using LLMs to continuously scan software updates, spot newly fixed bugs, and build attacks before teams can apply the fix — all without human involvement. That’s according to the latest report from the Google Threat Intelligence Group. We're helping defenders counter automated threats with Google AI Threat Defense, which combines Gemini and other frontier models to find and fix flaws before attackers can exploit them.
7
5
1
18
2,497
Along with some more recent thinking on cybersecurity defense factories
this week felt like sci-fi 👀 agents talking across isolated VMs through CPU cache timing, hardware bugs AI can exploit directly, lifelike voice models running on a laptop is patching enough anymore? 🤔 sam altman doesn't think so, jason clinton says hold on for 18 months, matthew prince says years of sky-is-falling weeks black october is real and we have to build the cyber defense factory sooner rather than later, agents running 24/7 and fighting compute with compute 🔥 Here's one take on what this could look like - notice that it will take a village to make this work! Lots of opportunity to build and reimagine and reinvent but the time was yesterday! full post here: whatshot.vc/p/whats-in-aiinf…
1
207
Did a deeper dive back in August post Blackhat on my weekly newsletter
What's 🔥 #510 dropped - post Black Hat 🚨 The agent swarm future is already here. Most of cybersecurity doesn’t see it yet. The OpenAI/Hugging Face incident made one thing clear: attackers will scale with tokens and compute. Defense must do the same. Compute fighting compute 👇 Full post here: whatshot.vc/p/whats-in-aiinf…
1
91
Nobody starts a company to negotiate the same SaaS contract for the 50th time. Common Paper's Cloud Service Agreement v3 is out. 10,000+ companies have signed on Common Paper, and these standards close hundreds of millions in deals every year. Less redlining, more closing
Common Paper Cloud Service Agreement v3 is out. It’s our most-used standard, and 10,000+ companies have signed on Common Paper. Every version is shaped by those real deals and by our committee of attorneys. Thank you to them.
1
2
351
Don't over-buy HBM, tier it. @netpreme's CTO showed that a separate memory tier for the KV cache delivered about 3x the batch size and 2.4x the peak decode throughput per GPU on GLM-5.3 at 128K context. Memory is the new inference bottleneck. teapawt.org/netpreme/status/212149…
“Don’t Over-Buy HBM, Tier It!” In our GLM-5.3 study at 128K context, the modeled memory tier supported ~3× the maximum batch size and 2.4× the peak decode throughput per GPU vs. the PCIe host-memory baseline. Our Co-Founder & CTO Nikita Lazarev shared the study at SGLang Summit today. HiSparse retains the full KV cache. A hot buffer stays in GPU HBM; the full cache can live in an external memory tier. Full talk video Monday, November 16. Thanks for having us, @sgl_project. #SGLangSummit #LLMInference
4
1
4
532
we need much more of this - pro founders and growth 📈
2
4
39
3,140
Nvidia fallout... Will all roads lead to the largest labs? My fear on a potential Open Weight Squeeze A risk, not a prediction. What’s 🔥 👇 #524
3
3
2
13
4,726
and more of this from everyone! teapawt.org/ClementDelangue/status…
We’re not going to win by lobbying, plotting, or manipulating. We’re going to win by building, in the open, with the whole community! Today, with @ggerganov we’re super excited to announce the release of llama.cpp v1, the first major fully stable version of the most used library to run AI models locally, meaning on your phone, your laptop, and your workstation. It means you can own your intelligence for free, not send your data to anyone and be 100% in control, wherever you are in the world. It shows our commitment to support every hardware, platform and model type out there and makes AI distributed rather than concentrated in a few Silicon Valley orgs. Huge thanks to the team, the long-time maintainers and every contributor who helped make this happen. Let’s build! 💛🧡💚
2
450