I'd look at rng initialisation first.
7,319 karma · joined September 21, 2012
I'd look at rng initialisation first.
IMHO the main thing thats interesting about AI assisted bug hunting is that it changes the balance of power from people who had a lot of free time & attention to the state and big business, who have money and frontier model access. It's a broadly "conservative" development in the sense that it distributes more power to groups who've already got it.
Waiting for the cyber "proxy wars" where state A equips deniable groups x, y with frontier access to undermine state B.
For up to low hundreds of users I think you're better off just with 1 vertically scalable box for all the officey / web server workloads.
You mitigate the hardware failure stuff with a vendor contract where you can get someone on-site and overnight you parts, and by keeping things super boring. Volume replication is not boring, avoid at all costs. NAS or SAN if you have to but all disks in the main box for as long as you can.
For 20 person SME maybe a 2-bay Synology or similar, for a heavier company a low end 2U with hardware support. Proxmox under the OS for reduced worry snapshots, rollback, backup etc. Proxmox is there for operational flexibility, resist the temptation to create a network of VMs, you just need 1 CT or VM with all the workload inside it.
For container workloads on 1 host Portainer works as well as k8s IMHO, it gives you the key property you want - you can IaC everything declaratively with terraform + compose over an API.
Caveat that if CI gets heavy you might need to scale that out but you can keep it stateless.
Social media "regressed" from the point of view of users because the success metric from the network's point of view was value extraction per eyeball-minute. As long as there continue to be strong financial incentives to have the strongest coding model I think we'll see progress.
Every day I deal with bad judgement calls from humans (sometimes my own!), but I don't screenshot them because it's not polite.
I don't think we're at the top of the curve yet? Current AIs have only been able to write code _at all_ for less than 5 years.
Code in particular is a domain that should be reasonably amenable to RL, so I don't think there are any particular reasons why performance should top out at human levels or be limited by training data.
The way I'm thinking about it is, it's a _workflow_ innovation?
So you ask for data sheets for all the visible chips and get PDFs in an output directory with minimal user interaction except to flip the board, ask for a basic idea of connectivity, get a stitched high res surface image etc.... which of course are all currently possible, but you can do them potentially with very low effort. There doesn't have to be a _software stack_ ahead of time. You ask Claude to do the thing, it will figure out how to do it, write some code, pull in some OSS and make the thing happen. You can take this project's software or leave it.
You might say "tell me where you think the JTAG headers are" and it will come up with a workflow to do its best at that task (most likely with variable results...), but nonetheless this is not a thing you can ask of any commercial product I am aware of today. With probes, stuff can get interesting.
Of course experienced hardware & reverse engineers already can do all this stuff and have a plethora of workflows for it but I still think it's an interesting POC of a generalisable approach. You can take or leave this particular software stack. Also, the hardware barely matters, you can duct tape whatever to whatever.
You could make a car that's safer than others at 10x the price but what would the demand look like at that price?
Would you pay 2x for your favourite software and forego some of the more complex features to get a version with half the security issues?
Where I DO see MCPs getting actual use is when the auth story for something (looking at you slack, gmail, etc) is so gimped out that basically, regular people can't access data via CLI in any sane or reasonable way. You have to do an oauth dance involving app approvals that are specifically designed to create a walled garden of "blessed" integrations.
The MCP provider then helpfully pays the integration tax for you (how generous!) while ensuring you can't do inconvenient things like say, bulk exporting your own data.
As far as I can tell, that's the _actual_ sweet spot for MCPs. They're sort of a technology of control, providing you limited access to your own data, without letting you do arbitrary compute.
I understand this can be considered a feature if you're on the other side of the walled garden, or you're interested in certain kinds of enterprise control. As a programmer however I prefer working in open ecosystems where code isn't restricted because it's inconvenient to someone's business model.
So I guess you had to modify the agent harness to do this? or I guess you could use... mcp-cli ... ??
NoVa literally one of the richest regions on the planet and it's all sorta tied up in the same thing. Seems unfair to say DCs "bring nothing", the whole ecosystem is a manifestation of concentrated defense spend.
Consequently, you wouldn't necessarily expect ancestral "defaults" to be optimal for modern environments.
The expensive equipment is usually a small (but crucial!) part of research activity, which involves things like talking to a lot of people, getting permission to do weird or new things, going out into the environment and collecting things in very specific ways, storing and transporting them carefully, observing, etc. Building or modifying existing lab instruments, doing various things with animals that are not co-operative ... and CLEANING. Who does all the cleaning?
Definitely use cases when you have a specific protocol you want to scale, but I'm also not sure how safe I would feel around AI with a license to experiment and access to dangerous reagents, high temperatures, etc. Or, god help us, an oligonucleotide synthesizer. Which is definitely going to happen (if it has not already).
Also worth noting that Birren was a paid consultant for DuPont - the company that made the paint. From that perspective alone you would kinda expect he is gonna pick colours they already produce reliably at scale.
He also consulted for army, coast guard and navy so whatever colours you pick for hazard marking in an industrial setting have to be _vaguely_ congruent or you're going to cause accidents.
Also by 1944 there would have been a ready made supply chain due to demand from the navy, which would have picked it for similar reasons and consumed it in enormous volume.
I think, practically, control rooms are chrome oxide green (you get to add as much titanium as you like - thats cheap as dirt too EDIT: it would have been lead actually in the 40s) for much the same reason that barns are red.
So it seems like it's a difficulty classifier for task descriptions written in English.
This is then used to score embeddings of Python code, which is a completely different distribution.
Presumably it's going to look at a simple solution, figure out it lands kinda close to simple problems in embedding space and pass it.
But none of this helps you solve harder problems, or distinguish between a simple solution which is wrong, and a more complex solution which is correct.
Parents want to signal "this child is looked after and we have a lot of capacity". Clothes, lunches, a lot of things are quietly like this.
A good hand-made gift demonstrates the status of the giver and provides proof-of-work for their regard of the recipient.
A 24 hour wait like this can sometimes be the result of a security team not knowing what else to do. There are all sorts of weird threat models when you think hard about how devices are used, like partners who have legit access to a phone at a certain point in time.
Also more mature businesses are different from startups. A mature business would not know what to do with "100 random little ai-powered mission-aligned startups under your control.", in fact that typically reads to management as chaos.
I mean, people who will create negative utility in a workplace or cost more in supervision expense than you get from them as output.
They create hazards for others by being drunk or on drugs on the job, or by harassing or bullying others, injuring themselves or others, causing personality conflicts or dramas due to trauma or unresolved mental health issues etc. I don't mean this as a value judgment, it's just like some people really aren't in a place in life where they can function well in work settings.
I'm not sure how you "guarantee" something that is dependent on complex situational decisions.
In general, I don't find the argument that there were 38 million deaths from sanctions very convicing. That estimate is based on excess deaths and correlation. It is serious work. But it doesn't establish causal structure.
What I mean is, those studies don't have the power to distinguish between (regime -> sanctions -> deaths) vs (regime -> deaths + regime -> sanctions).
I would definitely agree that sanctions cost lives (e.g. unavailability of specialist medicines) but I don't see a specific mechanism whereby they cause "millions" of deaths without a complete failure of local governance.
Hi, I think millions is a drastic overstatement here which undermines the rest of your (often legitimate) claims.
Also Israel seems to have fairly normal relations with many countries in the region, the difference seems to be they are "countries not publicly calling for the destruction of Israel".
I would argue that it's _scaling_ software development?
Making it so that you can create bigger programs: that accomodate more contributors, more requirements, faster, with more features, while being more correct, etc, without everything collapsing into spaghetti and chaos?
Historically, keeping a close eye on lines of code spent helped a lot along most axes. Finding ways to keep "code spent" down will remain very valuable but I don't think it's the _only_ way to scale?
Along the way we've developed architecture conventions, useful interfaces, test practices, standards, protocols, plugins & extension points, package repositories, etc. There are also social technologies like mailing lists, SIGs, bdfls, modern project & product management, etc etc.
To the extent that simplicity is measurable (less LoC spent, less code coupling, less cyclomatic complexity) it seems like it can be addressed as an optimisation problem and automated.