ghm2180

Whoa. This is amazing. I think What is interesting is whats driving this: the psychologically formed habit of asking questions to bots rather than search engines.

If the median query on the internet is now and forever answered by an AI bot, humans could just as easily become really important as an "authentic" source of answers.

I assume you run a benchmark on all the tasks on every change the model recommends? Do you have it set up like nanogpt speed run thingy that karpathy for a thumbs up or down for every change when you did this with astra?

There is already a big market for open model hosting and the accompanying demand for people and DCs. If you know how to deploy models — yes even open ones — and scale inference you are a hot engineer in demand.

Thank you for making this. I would like to make one simple probably obvious observation. You can watch this long time and not feel shitty about people or with conflicting thoughts racing in your subconscious the way you do with youtube.

I think the only way MS knows how to make money is lock someone into an ecosystem. I think their next attempt should be a closed source lean checker that has will have proprietary code for NL -> lean formalization checker.

I evolved my mental model for adhd by reading newer research like:https://www.economist.com/leaders/2024/10/30/adhd-should-not...

Metaphorically and analogy wise: If you think as genes as the holes on a flute, gene expression as the music then environment is what determines "bad" music. But there is no one specific gene that is known or corellated specifically for adhd symptoms.

ADHD is likely genetically caused but the genes are in everyone just expressed in specific environmental circumstances impossible to know for sure in a lab(cause they mostly look at outward behavior in lab settings and fmri blood flow).

But I've been doing meditation(45mins 3-4 days a week) I've also learnt for myself that 30 mins of power naps does wonders for my mind. The medication's effectiveness AND gated on the former two; SNRI, Dopamine reuptake inhibition is like carpet bombing and only makes you ready for rewiring: you still need to rewire with therapeutic and Behavior Therapy techniques.

But I know for a fact that my brain waves are different on days I do meditation and power naps.

I really do want to know if deep brain neural activity detection data can detect bad waves and possibly close the loop by providing corrective waves.

Noob q: Does this(and most) hacks like this on the iphone likely rely on some kernel module — or what it talks to for executing code — being compromised to run arbitrary code execution? I dunno about ios but it seems from the videos that something that reads imessage attachments was compromised for privilege escalation.

Like what im mean is an attack vector shaped similar to https://pepsipu.com/blog/2026-10-switch2-jailbreak-ns2jb/ i.e. they compromised `fs` module that did not protect against a buffer attack(in memmov)? My understanding is that most core kernel modules are difficult to attack directly because of XOM, PAC and CFI like protections make it impossible for anyone/thing, so most likely only indirect attacks like above can work in such situations — there is a lot of incentive for teams like apple to do everything possible to improve security so i assume only these kind of difficult to find vectors remain.

I think its largely because there is not a lot of money to be made by doing that in the short term — core non software services are still the largest part of the GDP(think GM, coca cola or united healthcare) ? The reasons are a bit varied but I think a neat abstraction comes down to incentives.

There is a reason why healthcare and grade school education are so hard to disrupt[2] (in the sense of United Health Care or wily publishing disrupted). The biggest incentive is cost and there is no reason to believe things will become less expensive unless people believe that — en masse. People only adopted the light bulb decades after its invention because industrial use made electricity dirt cheaper than wax and gas(and infra became ubiquitous).

Are AI tokens are making health care any cheaper, education any cheaper? No way. Do they make up majority of the economy? yes.

Also, here be dragons(no pun intended): Even china with a highly vertically integrated[1] system is facing so many issues with all incentives it is giving to AI companies, internally in china demand has collapsed, their export driven economy is not really sustainable to grow GDP at the +ve rate they want.

[1] https://www.economist.com/china/2026/03/24/chinas-government...

[2] When i worked for an ed-tech startup there is no incentive strong enough to use AI/ML, a big publishing exec told us "I don't need my product to have AI to succeed, I just need to say it does". There are however plenty of other "side incentives" like Google Chromebooks, Charter schools, but none for teachers and parents to replace Big Publishing.

I've been thinking a lot about this problem. Also, I think like the idea might be to look at the issue with code mode. Like basically, we know that all agents can write bash and execute it and that is what agents default code mode is.

But like if you only allowed for example JavaScript code that could be executed in a WASM sandbox, then you would have better control instead of just locking down with proxies. Basically there would be a very large API, basically tools, that the agent would codemode execute. My guess is that this would need a tool router to optimize the loading of these into the context. But it would give you very fine control over what code execution can and cannot be done.

Russian network ingress/egress?

> It is a research bench, not a ROM: a session lasts a given number of seconds, then the kernel returns to the loader and hands over its log.

I am "disappointed". no AI OS in the boot sector? It should be taking That log and AFKing the next fix man. Also "sorely missing" is a motorized rig controlling any hardkey presses for reboot and volume etc, because its not real AFK without that.

CmmonCommon 2026!

One small suggestion: Make the code open to the right instead of a new window, that makes it easy to keep context without distraction.

_Most_ people park liquid money in mainstream bank's checking accounts. This money is FDIC insured(upto 250K). The job of this bank is to is money is used by banks to make loans (duh!). There are three "modes" (in increasing yield_ order offered by these banks to account holders:

1. [FDIC] Put it in a Chase, BofA account. < 1% typically. Abysmal returns.

2. [FDIC] A higher yield checking/saving account(Ally at ~3.1 is a good example): But more hoops[1] to get more than 3.5%(today). Here https://www.wsj.com/buyside/personal-finance/banking/best-hi... the list.

3. [Not FDIC but pretty good**] A higher yield Money market mutual fund. You still need to place a buy order[2]. Something you have to do manually can give you around 3.86% today(See SNAXX)

This last case is more _interesting_ because its easy for agents. But consider what happens when a lot of people use muse to enmasse do this. There can be a run on the banks because[3] its not the same "money" in the balance sheet of the bank. Money leaving checking/savings and entering a MMF is not on the balance sheet of the bank(even though it can be coincidentally the same institution e.g. schwab checkig and schwab brokerage). Agents are a BIG threat to regular banks. I don't even see how it can be prevented, local agents are ubiquitous now.

[1] Exceptions are HY checkings but its hard to chase yield logistically: more hoops to jump like minimum balance or max ceiling, minimum transactions etc.

[2] See this from schwab's site: ``` For uninvested cash you do not plan to spend soon, you can consider investing in a money market fund such as the Schwab Prime Advantage Money Fund (SWVXX), which currently yields 3.71%, (7-day yield as of 2026-10-21), and can be purchased in a single trade with no minimum* and no transaction fee. Visit the Schwab Money Funds page to view all available funds and their current yields. ```

[3] https://www.youtube.com/watch?v=iPkJH6BT7dM

** MM funds are banks lending to each other to park overnight cash and borrow temporarily for actitivities, think of its like a bank's bank. In 2008 when the GFC was happening the fed had to step in and "ensure" this supply because a lot of banks were teetering and if one bank could not access its money or borrow for daily activities, no more cash would come from an atm.

I am surprised he is not wearing more protection. that thing looks like a hornet in bedroom test mode. Play safe.

I was wondering did he use the control mixer MCU (initially mentioned in the video) at all? The encoder should always feed into the MCU directly.

I learnt sheet music and played my first two piano pieces from scratch. It's the kind of meditative when you pay attention to learning little bits of it.

Very cool. I liked the high level overview where you can view it like do a funnel analysis for failures. Good for fast decision making.

I wonder if you have thought of using this for the voice AI space for creating traces from many test calls to speed up testing of voice bots? It's true there are a lot of challenges _before_ one gets to the stage of running actual calls like creating good test data etc, but I wonder if you see a definite role for this framework in that space?

I mean the people. I didn't mean or think that math research fundinng was a source of riches anyway; It's funded currently by mostly public money, how rich can any math researchers get from public money(compared to RSUs from stock grants)?

True. Pure Math does seem different. It's effects are felt way down the line and it's more about training people to think a certain way(what 3b1b creator pointed to). And I think when it's discussed they are talking about an entire field changing— material science, theoretical physics is still taught post bell.labs in unis — but I have to comceded I dunno if there is any part of pure Math that exists without the system of incentives, grant structures going into Universities that teach math...

Nice. I am curious if you would seek to maintain this with some kind of a foundation(apache) based funding or keep it internal to your work?

> I used AI a lot, but I was the one making the design decisions based on the tests that claude picked from my history of troubleshooting sessions as part of my job.

Cool, this seemed like a good learning experience, learning from testing it in a real world. I wonder if you plan to write up some of the experience in a blog or something?

Music is an excellent downtime.

It seems like part of the reason openai did what it did by not giving a heads up to the researchers about the release of the problems was that they might just want to divert resources away from math and become a math research entity unto themselves. The be all and end all of these research areas that AI is really good at. Like if you are beginning your math career you might just thing, why don't I just interview at openai, they seem to have a math "department"?

It's a bit crazy to imagine that world where frontier academic research as a publicly funded thing to train next generation could be dead and reborn in the hands of corporations.

HN profile.

The current cycle of AI and mathematics gives me great hope for ordinary people like myself to try and attempt learning some mathematics and try and see if I like the experience. Has any one tried that or is trying that? How are you approaching it? I would like to know - you can DM me - I would love to have a conversation if this excites you :)

I think the _key_ to knowing if I can do it has to come from the fact if my intuition about a problem helps me make progress. I would use AI to learn the parts I did not know and engage a community of people to ask questions. Both are key.

Also I am not 100% sure, but me beginning straightaway with lean to try something — like an engineer would — does not seem like the right approach before building some critical mass in the subject. It feels kinda wrong like building the cart before the wheel would.

Congrats on the launch. What's your turn around time to sign a BAA?

Again, its less work for people who can. Consider 3 something billion people in China and India. A conservative estimate is that at least 1/2 of them are living hand to mouth — could be tech workers, other service jobs, manufacturing jobs, farm jobs but all of these 1/2 are living paycheck to paycheck.

Most of these people cannot do most of these things, but it's not they can't or don't want to. The key IMHO is psychological/cognitive load to get past the day. A good example that OECD country readers can relate to is homeless people - an extreme but relatable example because you can see it. When you become homeless the first thing to go are dignity, self-confidence and in short order mind's ability to think beyond the short term(next 3 hours and today). And Living hand to mouth(just beyond the edge) is the same proverbial frog(the mind) in boiling water(circumstances), just slower and acceptable to one to not completely break one but deform the mind enough that it's not always possible to recover.

One abstract idea is to convert every digital event into a hook. The hook then goes into a local LLM that has a rule to figure out what to do with it. I use DDG or mozilla email aliases as much as I can. All my emails go to a local email client and for the non-spam classified mails I have a few rules(**) I am toying with right now:

1. If non-money sub, categorize its content and create a P1 watch event and a set of follow up actions.

2. If its a money/trial sub, create a P0 watch event and its own set of followup actions.

All watch events are queued for me by priority where i ask on the weekend: "Tell me P0 events", it then asks me what action. The ergonomic factor here is I the human can tune this valve on the information hose. There is also a dashboard for P0 notifications and an android push notification that notifies me of P0 events, but they aren't "consumed" until i decide. There are tons of places this can be useful: make progress on non-work related todos, systematic investment, writing year-end reviews.

None of this is rocket science.

**Local service webhooks are the key: Every web-hook has some kind of decision maker in it - an LLM + deterministic logic.

What's weird about this article is none of it is technically wrong: "Become a cheetah(super at ___)", "Aspire and engage to be better", "Build 6 types of capital". They are not bad things to do, if you can: and that is what it misses, if you can.

But don't feel bad if you have not been able to do any of them, e.g. Most People with kids can often do very little than job and take care of kids and that's it. What is more important is to not beat yourself up about it, give yourself grace and do one or two things that make you and your near and dear people/pets happy.

Thanks. I am curious were you trying something like pipecat's Turn finalization? https://docs.pipecat.ai/api-reference/server/utilities/turn-...

Pipecat already does this. It does it for end of turn detection and it's pretty damn good. It's just that it calls the same LLM I have to do it, potentially expensive and prone to latency and not calibrated.