CTO, Co-founder at https://nitter.gq/t.co/5v7vwhRvWY | Ex Deepmind, Amazon, JHU PhD, IITD ECE
Palo Alto
Joined May 2012
- Tweets358
- Following849
- Followers608
- Likes58
Pinned Tweet
We have launched the Harden AI Firewall today! Get it for free at harden.run/#firewall-install to safeguard your coding agents from data loss, secret exfiltration, and destructive actions.
64.63% with 1/2beta(50, 50)+1/2beta(0.5, 0.5) as a prior? Or simple 60% to entertain that coin maybe either fair or biased with no smoothing?
81
I see this sentiment in people that AI can't come up with new math because
> What's the objective function or RL env for "interesting conjecture"?
Interesting math was developed to solve practical problems, so the objective function for beautiful math is solving science.
1
1
65
> that AI disrupts society so much, or even kills vast numbers of us, that the preservation of something like the current mathematical tradition ceases to be of any concern: all that will matter is the survival of the human race.
I've written a blog post responding to the letter about maths and AI signed by 25 Fields medallists. As with the Leiden Declaration, I didn't sign it, but I agree with much of it and welcome its existence.
gowers.wordpress.com/2026/09…
184
Pushpendre Rastogi retweeted
Protection is foundational in financial services.
Harden is proud to join the Financial Institution Insurance Council, FIIC.
As AI changes how financial institutions build and operate software, security and control need to be built in from the start.
Learn more about FIIC: fiinsurancecouncil.com/
Pushpendre Rastogi retweeted
Introducing Harden Agentic Integrity Foundation (AIF), the unified security layer for AI coding agents
We believe the future of software engineering is long running agents taking on complex work while keeping your data, credentials, and critical secrets safe.
So we built AIF
→ Evaluates every action before execution
→ Uses developer intent and session context
→ Blocks unsafe actions or makes them safe
→ Runs locally and records every decision for audit
Whether you use Claude Code, Codex, Cursor, Antigravity, Kiro, OpenClaw, or Hermes, AIF works with the agents you already use.
AIF is not another dashboard or copilot. It is the security layer between your coding agent and the systems it can affect
Our post-trained model outperforms frontier models across critical agent-security benchmarks
AIF is free for individual developers: harden.run/?utm_source=twitt…
Need help getting started? Book a call, and our team would be more than happy to help you get set up: calendar.app.google/GdL3BvBu…
Latest Newsletter harden.beehiiv.com/p/a-big-w…
We just published our latest newsletter, and it captures one of the biggest weeks we have had as a team.
Inside this edition:
→ Harden AI Firewall is now live and free for individual developers. It watches every action a coding agent takes and stops the risky ones before they run.
→ We gave the AI Scanner a full redesign, with the focus on user experience. Applications, repos, scans, findings, and fixes now live in one clean workspace, so you can actually see what needs attention instead of hunting for it.
→ A recap of our week at Ai4 in Las Vegas, meeting developers and security leaders who are looking for a better way to run agents safely.
We have been pushing ourselves hard to get here, and there is a lot more coming.
Follow @hardenrun and subscribe to our newsletter to follow along as we build: harden.beehiiv.com/p/a-big-w…
2
259
First reaction is that this is so cringe, such a dumb take. Can this person actually be this stupid? but then I think maybe he is playing 4d chess? Maybe the point is to be controversial, and thereby retweeted, similar to what slate star talked about. In that case, 10 points for effort?
If you think about it, it's pretty embarrassing that frontier labs still pretrain models. Don't they know there's a theory for this? Given how big the model is, and how long you're going to train, you can predict the loss exactly. No need to actually train. They spend billions on this!
294
How I learned to stop worrying and love the open weights.
youtu.be/73rG0zsKoaA?si=zl--… via @YouTube
1
267
Pushpendre Rastogi retweeted
Excellent talk by @thjashin:
- What separates autoregression and diffusion?
- Bridging the gap between continuous and masked discrete diffusion via loss reweighting
- Insertion-based sequence generation
Very clear and easy to follow! (via @mblondel_ml)
youtube.com/watch?v=kMimQxIJ…
Amazing talk by @shmatikov at the real world AI safety conference on misaligned agents without any prompt injection.
2
375
I am hiring a founding security engineer for harden.run. harden.run/careers/founding-… . DM me for details.
- We are focused on AI DevSecOps
- We have hired a senior engg. with 20YOE in this area
- Our paper on prompt-optimization was accepted at ICLR and is out on ArXiv.
- We are on track to cross the million dollar ARR threshold in a month
- We have grown from 2 to 6 already and on track to add two FDEs soon
3
286
GPT5.4 is much better than Opus for high effort thinking.
GPT-5.4 Pro solves Erdős Problem #1196!
Very pleased with this result; definitely my favourite thus far! This problem has been thought about for some time which makes this reasonably impressive and meaningful (see Lichtman's comments below).
Formalisation is underway!
1
221
GEPA asks 'why did this fail?' — but the signal is still a score. The jump is using contrastive pairs: show the model the failure case *and* the success case side by side. It stops guessing at failure modes and starts extracting structural rules.
This is why ContraPrompt gets +29% HotPotQA vs GEPA. Same data, different feedback format.
vizops.ai/blog.html
120
Analog IC layout is one of the hardest AI benchmarks: spatial reasoning, multi-objective tradeoffs (matching, parasitics, routing), no automated P&R tools.
We ran VizPy's ContraPrompt on it. The optimizer mines failure→success pairs across iterations, extracting layout strategy rules the LLM learns to apply.
Result: 97% of expert placement quality. Outperforms RL fine-tuning of a 120B model by 26%. No domain-specific training data.
vizops.ai/blog/prompt-optimi…
1
113
The optimizer discovers layout strategy from scratch — no hints. ContraPrompt extracts: 'separate PMOS/NMOS into distinct columns, align drain-paired devices.' These rules encode *strategy*, not coordinates. The LLM fills in circuit-specific values at generation time.
TTT (RL fine-tuning, 120B model) memorized training circuits but scored 0.502 on test. Prompt optimization scored 0.634.
51
We gave an LLM a 9-line random-search stub and a blackbox objective. 5 rounds of contrastive feedback later, it writes a solver that beats Optuna on 96% of benchmarks (53/55 EvalSet problems) at the same 2k eval budget.
The key: contrastive pairs surface landscape structure raw scores don't. The LLM learns geometry from paired failures vs successes — then rewrites its strategy each round.
Final code: 9 lines → 100-230 lines of specialized multi-phase optimization. No hand-tuning.
vizops.ai/blog/contraprompt-…
1
2
145