hazrmard

This is my takeaway:

Authors analyzed ~300 workers doing ~1000 tasks. They collected self-reported measure of using critical thinking [yes/no], and estimate of cognitive effort of doing the task. Authors also measured workers' confidence in themselves doing the task or the AI doing the task.

They found: if someone is confident that AI can do the task, they report using critical thinking less, even though they think the task is easy. When someone is confident in doing the task and in evaluating AI responses, they use critical thinking more, even though they think the task is harder.

I think the paper is saying: low-domain expertise (read: confidence) + unrestricted AI tooling can eliminate the human from the loop. The fallout is better automation or unreliable systems. We need to resolve this ambiguity.

The question is: should low-domain expertise be coupled with more restricted AI tooling as a critical thinking forcing function? Are schools/workplaces right in restricting AI so students/new workers can reach the point of high domain expertise?

A breakdown of why the awarded work, Ice Cube, is significant:

  - A neutrino is an elementary subatomic particle. Neutrinos are produced by nuclear reactions inside stars, supernovae, radioactive decay. They are one of the most abundant particles in the universe.
  - Neutrinos are known as "ghost particles". They have 0 charge and near-zero mass. They only react with the weak nuclear force and gravity. Incredibly hard to detect. Trillions can pass through a whole planet without hitting a single atom!
  - Neutrinos give us a pristine snapshot into the origin of the universe. They have travelled billions of years and trillions of miles without interacting with anything. Unless...we catch them!
  - Ice Cube does this. Located in Antarctica, the project turns a cubic kilometer of ice into a neutrino detector. How? Scientists drilled boreholes 2.5km deep into the ice and placed 5000 optical detectors to catch a neutrino interaction.
  - When a neutrino, rarely, collides with an atom, it produces charged particles. In ice - not vacuum! - certain particles can travel faster than light. This produces something similar to a sonic boom. A faint, blue glow known as Cherenkov radiation (see it in action https://youtube.com/watch?v=hSuSG19Pcoc).
  - IceCube was first to detect neutrinos coming from outside the solar system, establishing the source of high energy cosmic radiation. It also opened a whole new chapter of neutrino astronomy.

This is not intended as a product launch. I think this is a preemptive bet on a conjunction of the following happening in time for a future product launch:

  1. Cost to launch / mass lowering due to technological advancements.  
  2. Mass / compute lowering due to technological advancements.  
  3. Compute hardware evolution stagnating to the pace, relative to compute payloads, where CapEx is amortized over longer duration.  
  4. Compute & energy hardware reliability improving so OpEx is lower, and remaining useful life predictably keeps hardware running over its designed lifetime.

(I work on #4).

A wishy-washy inequality (per time, factoring out launch cost from OpEx):

    Revenue - CapEx/lifetime - (launch_cost + OpEx) * (1 + failures) > 0

Presumably, any one of these could see a x100 improvement and make the whole equation in the green. But, I believe this is a low-earth-orbit-shot bet on the sum getting better.

Note: I did 1+failures to denote baseline cost (initial launch, nominal OpEx).

I read this paper (https://arxiv.org/pdf/2410.13722) on pre-training poisoning of LLMs. The authors demonstrated that poisoning 0.1% of the pretraining dataset can measurably manipulate beliefs of the model.

What I read about is a more sophisticated supply chain attack. But these days, search previews use retrieval-augmented generation to parse websites to generate answers on the fly. In that case you don't need to poison the supply chain; good SEO is enough.

Turns out the balance of my PhD thesis was coming up with ways to cheer up disappointed drones >:(

I resonate with the author's point that I can modify my own website to my will.

I use hugo to build my website and I wrote my own theme. It's clunky, but control over the design/build lets me:

  - Put PyIodide into posts to make ML code explainers interactive (https://iahmed.me/post/poor_mans_autograd/)
  - Inject custom widgets to illustrate theoretical ideas (https://iahmed.me/post/surprise-derivation-entropy/)
  - Have multiple layouts to suit my needs (https://iahmed.me/resume/ rendered from json at build time for SEO)

On the flip side, I'm not good at SEO :) It works because SEO is not my primary goal, but I'd love to have a bigger audience. The flexibility I'm afforded is an ok compromise.

Using Django w/ frontend frameworks has been a struggle for me. (I only dabble as a hobbyist.) I think splitting the backend off into a restful API makes the marriage work. I've looked at django-rest-framework (https://www.django-rest-framework.org/) and django-ninja (https://django-ninja.dev/). Of course,, then you're discarding a lot of batteries that come with the django framework. Perhaps using Django templates for admin / internal facing work only, and React for customer-facing sites can be a healthy compromise?

Does the peak starting in mid-2020 correspond to mass speculative / adversarial / bandwagon hiring by tech companies during the pandemic?

Looks fun, but I don't quite understand this product:

  - Do the buttons map to configurable skills / prompts?
  - Is it meant to be used remotely with some independence (like codex remote), or is it a peripheral like a trackpad?

You'd be right to be surprised by lack of mention of quaternions. I (blog author) was not too familiar with them when I did the work this post is based on. With the benefit of hindsight, I may yet revise the post.

Hello, blog author here. I agree the math appears overkill. I wrote this as a learning aid for myself with the benefit of hindsight of having worked with some drone sims. I wanted to dispel any doubt in my mind that I could derive drone physics from basic principles, instead of copy-pasting state equations. I went further and tried to motivate transport theorem, rotational analog of F=ma etc from scratch.

In summary: I take F=ma and extend it for rotational motion. (1) Calculating linear motion when the vehicle containing sensors is rotating. (2) Calculating rotation of the vehicle itself due to thrust/yaw force acting about its center of mass.

I'll echo what the other commenter said: this is no way PhD math. It may appear so - but I'm only being verbose with simpler concepts like cross products and rotation matrices.

It's my blog. Bartosz Ciechanowski's work was my inspiration!

Good work! I wonder if meta-learning can play a better role here compared to heuristics or hindsight. MAML requires hessians, but first-order MAML or Reptile variants could help apply layer-wise adjustments to learning rates.

These day's I'm super-into information theory and entropy, so I liked the connection made at the end. I'm a visual person, and I found this schematic of SVD on the wikipedia page very insightful [1].

[1]: https://en.wikipedia.org/wiki/Singular_value_decomposition#/...

This is very impressive! I researched fault-tolerant octorotor control using RL in grad school for a NASA project. Perhaps this may be helpful[1, see section 8.3]! The field is moving fast, so there may be better or more suitable approaches out there now.

For folks who are interested in UAV physics, I wrote up an explainer[2].

[1]: https://drive.google.com/file/d/1RTEVRd0XCWLuDXY2nkbmYuOaa5x...

[2]: https://iahmed.me/post/drone-physics/

Check my understanding & follow-up Qs:

An auto-encoder is trained on [activation] -AV-> [text] -AR-> [activation], where [activation] belongs to one layer in the LLM model M.

Architecture.:

    Model being analyzed (M): >|||||>  
    Auto-Verbalizer (AV) same as M, with tokens for activation: >|||||>  
    Auto-Reconstructor (AR) truncated up to the layer being analyzed: ||>

The AV, AR models are initialized using supervised learning on a summarization task. The assumption being that model thoughts are similar to context summary.

The AR is trained on a simple reconstruction loss.

The AV is trained using an RL objective of reconstruction loss with a KL penalty to keep the verbalizations similar to the initial weights (to maintain linguistic fluency).

- Authors acknowledge, and expect, confabulations in verbalizations: factually incorrect or unsubstantiated statements. But, the internal thought we seek is itself, by definition, unsubstantiated. How can we tell if it is not duplicitous?

- They test this on a layer 2/3 deep into the models. I wonder how shallow and deep abstractions affect thought verbalization?

Very cool! Interesting how the choice of solver affects the solution. Euler doesn't handle misbehaved equations very well. You can see this in the Helix setup where the bodies just fly off.

I should read up on Tailscale more. I have been using ddclient[1] or the router's built-in dynamic DNS[2] to set up my servers / homelab. This leaves the endpoints exposed to the public internet, as the author says.

    [1]: https://github.com/ddclient/ddclient  
    [2]: https://kb.netgear.com/1058/What-is-Dynamic-DNS-DDNS

cue the bell curve meme for learning autonomy:

                 ____.----.____
          ______/              \______
    _____/                            \_____
    ________________________________________

    (simulations)  (real world data)  (simulations)

Seems like it, no?

We started with physics-based simulators for training policies. Then put them in the real world using modular perception/prediction/planning systems. Once enough data was collected, we went back to making simulators. This time, they're physics "informed" deep learning models.

    https://iahmed.me

Hugo website, with a theme I made from scratch myself.

Github Pages deployment.

Here's my first website from when I was in college and had no experience in web dev. I still keep it on for nostalgia:

    https://iahmed.me/old_www/

This reflects my experience. Yet, I feel that getting reliability out of LLM calls with a while-loop harness is elusive.

For example

- how can I reliably have a decision block to end the loop (or keep it running)?

- how can I reliably call tools with the right schema?

- how can I reliably summarize context / excise noise from the conversation?

Perhaps, as the models get better, they'll approach some threshold where my worries just go away. However, I can't quantify that threshold myself and that leaves a cloud of uncertainty hanging over any agentic loops I build.

Perhaps I should accept that it's a feature and not a bug? :)

The paper finds:

- For LLM-assisted output, the more complex the LLM-writing is, the less likely the paper is to be published. From eyeballing, at WC=-30, both have similar chances of publication (~46%). At the upper range of WC=25, LLM-assisted papers are ~17% less likely to be published.

- LLM-assisted authors produced more preprints (+36%).

I wonder:

- What is the distribution of writing complexity?

  * Does the 17% publication deficit at WC=25 correspond to 17% of the 36% excess LLM-assisted papers being WC=25, thus nullifying the effect? Although, it puts extra strain on the review process.

I can vouch for this with my experience.

Back in grad school, I was out making new friends. I was playing tennis 4-5 times a week. I'd invite players for post-game coffees (in the morning) and dinner (evenings) at every game. Consistency mattered. I'd ask every time. Slowly we had our regulars. Our coffee times became an institution in and of themselves.

People are busy, yes. But, people also want to be in demand. People also don't want to be rejected. And, people also don't want to be left out.

Asking around, I was exposing myself to rejection. Some folks appreciated their time being demanded. More still joined because they didn't want to be left out.

Do I understand this right?

A light-weight speculative model adapts to usage, keeping the acceptance rate for the static heavy-weight model within acceptable bounds.

Do they adapt with LoRAs?

This takes me down a memory lane! For my undergrad capstone project, we made a cubesat tracker for our university's satellite using a RPi/Arduino/Software-defined-radio to receive transmissions every time it passed over us. I cringe a little looking at the code now - but it worked!

I agree, cubsats are a wonderful way, for college students even, to tinker with space(-adjacent) tech.

https://github.com/hazrmard/SatTrack

I am working on a budgeting app!

Features:

  - Local. No internet connection needed.  
  - Manual. Every transaction is added by the user.
  - One-off or arbitrarily recurring transactions.  
  - No lock-in. Check out your data any time.  
  - Arbitrary metrics to track performance. 
  - Hosting on the cloud for mobile access. 

Why?

I've been using Google sheets + forms for the last 8 years to track my finances. It's worked well, except for minor inconveniences. This app is my answer to my own problems.

Thank you very much! I had been reading up on effect of diet for treatment & management. I like the focus on citations.