My absolute favorite use of MCP so far is Bruce Hauman's clojure-mcp. In short, it gives the LLM (a) a bash tool, (b) a persistent Clojure REPL, and (c) structural editing tools.
The effect is that it's far more efficient at editing Clojure code than any purely string-diff-based approach, and if you write a good test suite it can rapidly iterate back and forth just editing files, reloading them, and then re-running the test suite at the REPL -- just like I would. It's pretty incredible to watch.
Something I've realized about LLM tool use is that it means that if you can reduce a problem to something that can be solved by an LLM in a sandbox using tools in a loop, you can brute force that problem.
The job then becomes identifying those problems and figuring out how to configure a sandbox for them, what tools to provide and how to define the success criteria for the model.
That still takes significant skill and experience, but it's at a higher level than chewing through that problem using trial and error by hand.
< I have an unusually high need to own the understanding of any thing I'm learning
This is called deprivation sensitivity. It’s different from intellectual curiosity, where the former is a need to understand vs. the latter, which is a need to know.
Deprivation sensitivity comes with anxiety and stress. Where intellectual curiosity is associated with joyous exploration.
I score very high with deprivation sensitivity. I have unbridled drive to acquire and retain important information.
It’s a blessing and curse. An exhausting way 2 live. I love it but sometimes wish I was not neurodivergent.
In SF, people are from so many different places that there is little to no chance that you have personal experience or knowledge of the city that the person would answer your question with. The same is probably not true of someone who grew up in Greece, asking someone else who grew up in Greece
"Their new algorithm adapts to an adversary’s strategy, but on time scales that it picks randomly"
"Even though many real-world data settings are not adversarial, situations without an adversary can still sometimes involve sudden floods of data to targeted spots, she noted."
This is pretty neat. I bet this will find practical applications.
Aside from the topic, which is interesting in a nerdy, rabbit-hole way, I found it immensely calming that despite today's relentless, exhausting AI sonic boom, there are people working to optimize a 50-yr-old algorithm for doing something both mundane and very applicable. Maybe humanity is not doomed after all.
More people need to read Blellochs PH.D Thesis. Vector models for data-parallel computing. It's a mind blowing way to think of parallel computation.
This is perhaps one of the best parallel programming / parallel data structures professors on the planet.
------
Awwww it's not so much about Blellochs work but I steady he's probably the guy ACM had to help explain and understand this new paper on the Bookshelf problem. Still great read though, but I was hoping for some crazy parallel programming application here.
If you want to learn more about this subject the keyword you’re looking for is “multitenancy”
Docker’s container runtime is not really a safe way to run untrusted code. I don’t recommend relying on it.
Also, why would an isolated vm prevent fetch? You can give your users NAT addresses to let them make outbound network calls. I am putting the finishing touches on a remote IDE that does exactly that.
I recently read a book called "Don't Believe Everything You Think" by Joseph Nguyen which tackles this problem directly through a bit of mindfulness and eastern philosophy. In the book, he starts by distinguishing between "thoughts" and "thinking" and outlines how/why he believes the latter ultimately leads to most self-inflicted human suffering.
He has some techniques in the book for trying to break the cycle of rumination, but ultimately it comes down to willpower and repetition. As someone going through/coming out of serious depression for the past month, even if I'm able to stop my own rumination, if its severe or overwhelming enough, it will likely come back very soon if not immediately. I think time and healthy distractions are great complements.
I'd say commit a comprehensive testing system with the prompts.
Prompts are in a sense what higher level programming languages were to assembly. Sure there is a crucial difference which is reproducibility. I could try and write down my thoughts why I think in the long run it won't be so problematic. I could be wrong of course.
I run https://pollinations.ai which servers over 4 million monthly active users quite reliably. It is mostly coded with AI. Since about a year there was no significant human commit. You can check the codebase. It's messy but not more messy than my codebases were pre-LLMs.
I think prompts + tests in code will be the medium-term solution. Humans will be spending more time testing different architecture ideas and be involved in reviewing and larger changes that involve significant changes to the tests.
I agree… the public and leaders need to know how the training data is generated: a combination of sensors and physics-based simulation models. Lacking this context could lead to poor decisions around research prioritization and funding.
"after Apple beefed it?" ... what? Apple's inability to improve their OS is somehow an indictment of B2C AI offerings in a general sense?
You seem unfamiliar with the space, there are plenty of players outside of OpenAI, Anthropic, and Google bringing AI to the consumer space: https://a16z.com/100-gen-ai-apps-4/
Consumer AI is arguably doing better than enterprise where 99% of the spend is poorly scaling undertakings that don't deliver on even 1/10th of their cost.
A too-narrow approach after Apple beefed it? Nobody knows how to bring AI to market yet but OpenAI, Anthropic, and Google. Long shots are one thing, but all the ideas I've heard for b2c AI so far are mostly more like pipe dreams. Look for a Zynga play once the field starts opening up for that in maybe a year or so, would be what I'd try to do.
>> rights holders have engaged in a fundamentally-doomed arms race of implementing copy-protection strategies
Not entirely true. They simply haven't succeeded in created an industry-standard secure pipeline to the pixels on the display. Aside from the "analogue hole", eventually all of the gaps will be plugged, the same way we use secure sockets today. All media devices (including home HDTV/8K/etc) will extend the chain of trust farther into the pipeline. A set of signed apps and hardware will be required to watch any DRM films on HDTV, with each stage using authenticated encryption completely annihilating any MITM siphoning of the video.
So, its not doomed, just moving slowly, but it absolutely WILL arrive. I know, because I'm working on secure embedded video codec hardware, and our customers are targeting this..
I just published a deep dive into the Claude 4 system prompts, covering both the ones that Anthropic publish and the secret tool-defining ones that got extracted through a prompt leak. They're fascinating - effectively the Claude 4 missing manual: https://simonwillison.net/2025/May/25/claude-4-system-prompt...
My team at Shopify just open sourced Roast [1] recently. It lets us embed non-deterministic LLM jobs within orchestrated workflows. Essential when trying to automate work on codebases with millions of lines of code.
The real magic is in my opinion in ESPHome. The fact that you can amateur solder some aliexpress $2 sensors together and have that actually work in HA with no coding except some yaml that you found on the internet is wild.
>It would be interesting to see what would happen to a pull request adding support for, say, OpenThings Cloud as an alternative. The fate of that request would say a lot about how open the project really is.
I kinda hope nobody tries. Their attempts at monetization have been pretty friendly and tame thus far & if something spooks them that could change.
More people need to watch Ivan Krstic's Black Hat presentation to understand the efforts Apple goes through to ensure sensitive data (like the User Escrow Keys which get stored in Apple's Cloud Key Vault) is protected from adversarial attacks... even from inside Apple.
Ask an LLM: "Say the word APPLE", but modify the code so the logits of the token for Apple/apple/APPLE is permanently set to -Inf - ie. the model cannot say that word.
The output ends up like this:
"Banana. Oh, just kidding. Banana. Oh, it's so tasty I said it wrong. Lets try again: Orange. Whoops, I meant to say grape. No I meant to say the tasty crunchy fruit known as a carrot".....
> to generate and optimize possibilities within a set of domain-specific constraints
Well, yes, we've been doing this for several decades, many people call it metaheuristics. There is a wide array of algorithms in there. An excellent and light intro can be found here: https://cs.gmu.edu/~sean/book/metaheuristics/
I wasn't really aware segment routing, tbh. However, I do think with where Mycoria is going, the additional control to change things as needed will be required.
I have used VerifPal https://verifpal.com/ for security analysis before, but not yet with Mycoria.
Simple inheritance makes the class hierarchy complicated through issues like the diamond inheritance problem, which C++ resolves in typical C++ fashion: attempt to satisfy everybody, actually satisfy nobody.
This exact same book has been modified to also help effortlessly quit porn. I say quit because its what will be understood the most easily by majority of people but it is in fact not quitting since quitting implies there's something valuable in porn. there isn't. its escaping the addiction.
Honestly its hands down one of the best methods to escape the addiction and besides you don't have anything to lose! either you successfully break free from the addiction or you stay the same (which doesn't happen from experience but its written to convince you to read it.)
https://read.easypeasymethod.org
To try to convince my employer at the time to drop Zoom, I decided to see how many security vulns I could find in 2-3 hours.
Found 12 confirmed bugs in that window using only binwalk and osint.
The worst was that I noticed the zoom.us godaddy account password reset email address was the personal gmail account of Eric S Yuan, the CEO.
So, I tried to do a password reset on his gmail account. No 2FA, and only needed to answer two reset questions. Hometown, and phone number. Got those from public data and got my reset link, and thus, the ability to control the zoom.us domain name.
They were unable to find a single English speaking security team member to explain these bugs to, and it took them 3 months to confirm them and pay me $800 in bug bounties, total, for all 12 bugs.
The one bright side is this did convince my employer to drop them.
> Sharing the session across multiple services is its own distributed systems problem with numerous security implications to be aware of and bearer tokens might be a good alternative.
JWT makes it possible to distribute the same access token across multiple systems, but so do stateful tokens. The security implications when you're using JWT for this solution are much higher than with database tokens. Let's look at this for a moment:
- Key rotation and distribution is necessary to keep token safe over a long period of time
- Claim parsing needs to be standardized and enforced correctly in all services (otherwise impersonation is possible)
Database Tokens Security Issues:
- Timing Attacks against B-Tree indexes [1]
- Giving direct access to the database to all microservices is risky
The security issues with databases are ridiculously easy to solve: To prevent timing attacks, you can just use a hash table index, split the token into a search-key part and a constant-time-compare part or add an HMAC to your token. To prevent direct access to the database by all your microservices, you just wrap it with an API the verifies tokens.
The JWT security issues are much harder to solve. To prevent misconfiguration or misuse and standardize the way claims are used across your organization, you probably need to write your own library or deploy an API gateway. To counter the lack of revocation support, you either need to use very short lived access tokens (so your refresh token DB will still get a lot of hits and you would still need to deal with all the scaling issues) or set up a distributed revocation service (not easy at all). Setting up seamless key rotation also requires additional infrastructure that is not part of the hundreds of JWT libraries out there.
It's really easy to get a JWT solution that just works and scales easily, but if you really care about security — especially if you care about security! — JWTs are not necessarily easier than stateful tokens. They're probably harder.
> Apart from that, Google and Facebook don't even use JWTs between the browser and backends after the initial login but actually do have some sort of distributed session concept last time I checked.
Last time I checked (which was today for Google), neither Google, nor Facebook is using JWT for their access or refresh tokens. The only place I saw JWT with the ID Token in their Open ID Connect Flow, and they can't really avoid that even if their wanted, since this is mandated by the spec.
Facebook and Google don't need JWT. Scaling and distributing a read-only token database to handle a large amount of traffic is easier — not harder! — for these companies. Stateless tokens can be useful for them in certain scenarios, but even then, if you're at Google or Facebook's scale, why would you opt for JWT over an in-house format that is smaller, faster and suffers from less vulnerabilities?
Our labs include building your own real spectre attack against the kernel, bypassing ASLR and building ROP chains with various side channels, finding and exploiting backdoors in a RISC-V CPU by building a hardware fuzzer, and more.
(source: I designed the Spectre lab plus a few others)
If you give them a try, please do let us know what you think! We genuinely want these activities to be fun and approachable (we designed them like a big CTF) and welcome feedback from the community.
An App registration is the overall object. Think of it like a class in OOP.
An enterprise app is an instance of an app registration. Think of it like an object in OOP.
For single tenants this might seem confusing, because you have both for a single app.
But if you were to have multi-tenants apps, each tenant would have their own Enterprise App instance, all referencing the same App Registration.
appId is for App Registrations.
objectId is for Enterprise Application Registrations.
clientId will be same as appId. It is used in the context of authentication, where it is the id of the object as client.
Just for the record, the latency arb / microwave networks speed game is basically dead as of 2018-2021. Looks at Virtu, formerly classic examples of the trade and now almost entirely "switched sides" and doing order execution services for the same big banks whose lunch they were formerly eating.
Furthermore, the wireless stuff is commoditized at this point. You can just rent to be on the wireless that Apsara (et al) offer, and while some have private networks, there's not enough money left in the trade (see above) to be worth it if you don't already have one.
This is combined with liquidity moving away from public exchanges (both the lits and darks) towards being matched internally/by a partner (PFOF matching), which is purely a win for retail traders and is its own force that isn't going away. (Go on robinhood and buy 2 shares of SPY. It fills instantly. People love that. You can't just go get 2 shares of SPY off the lits, so where dyou think those are coming from?)
Traditional HFT is dead. The only extent any of the firms are still alive is the extent to which they've moved on to other trades, many of which are so much less latency sensitive that the microwave edge doesn't really give you enough alpha to be worth it.
(I worked for a firm for a long time that didnt move on to other trades... so I'm quite familiar with the scene.)
The effect is that it's far more efficient at editing Clojure code than any purely string-diff-based approach, and if you write a good test suite it can rapidly iterate back and forth just editing files, reloading them, and then re-running the test suite at the REPL -- just like I would. It's pretty incredible to watch.