←
ishaan81 posts
ishaan avatar

ishaan

@ishaans22
◷ Joined July 2020▧ 1 Photos and videos
Jai Kondapalli avatar
⟳ ishaan retweeted
Jai Kondapalli@jai_kondapalli·Oct 28

the homebrew cybernetics club of nyc™️ is convening for dinner Nov 4th will be discussing the Ant S-1, future of the labs as public co's, economics of AGI, and more! lineup includes an amazing set of folks from the labs, leading ai startups, simulation / prediction co's, ai rollups, and more we've saved a few spots for the timeline - dm if interested cc: @hypersoren @PossibltyResult @dbarabander @ishaans22

Post attachment
Mike Solana avatar
⟳ ishaan retweeted
Mike Solana@micsolana·Oct 16

okay one thing I can say with confidence is truly nobody thought the guy who owned the 140 character company would also be the guy that built the flying cars

Felix Stocker avatar
⟳ ishaan retweeted
Felix Stocker@felixstocker14·Sep 28

Today is a great day to start talking about Foresee: We’re building a better way to shop for home and auto insurance, designed for AI assistants and chat interfaces. Foresee puts State Farm, Progressive and 20+ other carriers side-by-side, allowing AI assistants like @Muse and Instinct to explore coverage options. Insurance is a commodity product, but big insurers use brand advertising to misdirect American consumers, preventing honest comparison. Meanwhile, a huge lead generation ecosystem makes money through misdirection, incomplete information, and spam calls. It’s a massive problem, caused by generations of perverse incentives. We think that information deserves to be free - and so we’re building tools for agents designed for maximum transparency. @foreseeinsure is now available for auto in California - and we’re hiring! go-foresee.com/careers

Post attachment
Rishi Desai avatar
⟳ ishaan retweeted
Rishi Desai@therishidesai·Jun 26

If you have a background in computer architecture or operating systems and want to work on real autonomous systems, my DMs are open.

The U.S. military confirmed that a Saronic Corsair was used in the first-of-its-kind rescue operation following the downing of a U.S. Army helicopter. At Saronic, we build autonomous vessels to extend capability into the most demanding and dangerous environments. Knowing Corsair could play a role in helping bring our service members home safely is exactly why we build. Proud to support the men and women of the U.S. armed forces — and grateful to the team that made it possible. 🇺🇸

Post attachment
Braeden Norris avatar
⟳ ishaan retweeted
Braeden Norris@braeden_norris·Jun 24

I wrote about GLM-5.2 as a window into frontier API economics. Near-frontier intelligence is getting cheaper and more commoditized, but frontier labs may still have meaningful inference margin on the last increment of quality and reliability. Link below.

Post attachment
underscore advait patel avatar
⟳ ishaan retweeted
underscore advait patel@_advaitpatel·May 14

Mind has raised an additional $400M, bringing our total funding to over $1 billion. We are building dexterous, general-purpose robots and foundation models for industrial deployment, starting with the automotive industry. I’m super excited and grateful to be working on these problems with such a stellar team. We are uniquely positioned to tackle the general-purpose robotics problem given our deep partnership with @Rivian, who is a shareholder and our pilot customer, and who is also providing us with data from production lines to add to our training mixture. It’s still super early days, and we’re hiring across research, software, hardware, and more! (Link below)

Post attachmentPost attachment
jaisel avatar
⟳ ishaan retweeted
jaisel@jaiselsingh·May 5

Let’s go! finally got post-physics parameter optimization working. Shared parameters across multiple real-world trajectories, fitting the sim closer to reality instead of tuning one rollout at a time :)

0:27
0:13
0:16

the gtsam tracker is essentially one shotting the Mujoco trajectory’s to match real! I still have to work on resolving mesh to mesh contacts for the objects and see what params actually carry enough information to be optimised for within the trajectory.

0:16
0:24
Joseph Jojoe avatar
⟳ ishaan retweeted
Joseph Jojoe@josephjojoe·Apr 24

I'm excited to open-source optimized MLX and Torch ports of the Chai-1 protein folding model! (1/5)

Post attachment
ishaan avatar
ishaan@ishaans22·Apr 21

Validated on four LoRA fine-tunes of Qwen2.5-0.5B: synthetic canaries, SST-2, IMDB, PubMedQA. On IMDB without DP, training examples have 32× lower loss than held-out and the attack hits 0.926 AUC. With DP at ε=9.7, it drops to 0.500 AUC for 0.6pp accuracy cost. The other three collapse similarly. github.com/ishaansingh22/pri… 6/6

ishaan avatar
ishaan@ishaans22·Apr 21

The full DP-SGD stack fits in ~600 lines: each step clips per-sample LoRA grads to L2 norm C, sums, adds Gaussian noise at σC, and averages; a pure-stdlib RDP accountant tracks ε online; `mx.compile` refreshes noise every step. 5/6

ishaan avatar
ishaan@ishaans22·Apr 21

Per-sample gradients are the hard part of DP-SGD: each example needs its own gradient so you can clip it individually. Opacus (PyTorch) registers hooks on every trainable layer to extract them. MLX's `grad` and `vmap` are function transforms that compose, so `mx.vmap(mx.grad(loss_fn))` is the per-sample gradient function. 4/6

ishaan avatar
ishaan@ishaans22·Apr 21

DP caps how much any single example can move the weights: no row leaves a fingerprint and ε formally bounds the leakage. 3/6

ishaan avatar
ishaan@ishaans22·Apr 21

private-mlx is impactful on tasks where you'd normally fine-tune on private emails or medical notes, ship it, and then a membership inference attack would recover training examples by flagging inputs with suspiciously low loss. 2/6

ishaan avatar
ishaan@ishaans22·Apr 21

I built differentially private fine-tuning for MLX so you can train local models on your private data. The full stack fits in ~600 lines and drops attacker recovery of training data from 90% to 50% 1/6

Post attachment