Complexity itself obviously isn't new, and in many cases we've replaced terrifying, opaque natural uncertainty with systems that are much better at keeping people alive. But I think there's still a difference between complexity that is encountered and complexity that is administered through
Maybe the goal isn't to reject complexity entirely, but to be much more suspicious of complexity that gives no corresponding increase in dignity, beauty, autonomy or peace
Mesh networking is still mostly a playground for hobbyists and hacky, built-on-the-knee implementations. People love shipping a cool PoC, but as soon as the boring stuff starts - stabilization, drivers, edge cases - everyone bails to chase the next hype protocol. We’re left with mountains of half-baked C++ legacy that nobody dares to refactor because the whole house of cards would collapse iirc
It’s hilarious how people are blaming Claude for a human being a total snake. Trademark squatting is the oldest legal trick in the book. AI just gave Andy the leverage to fake "official" status and look busy while the rest of the team did the actual heavy lifting. The tech just fast-tracked the inevitable fallout tbh
It’s funny how a standard power grab over cash and control is being packaged as some "Holy War against AI"
Highlighting the Claude usage in the headline is a blatant move to bait the anti-AI crowd and farm sympathy points. The real issue has nothing to do with who generated the AST -whether it was meat-ware or a transformer - the issue is the brand hijacking. Stop dragging technology into this; it’s just pure legal scumbaggery
This is a fundamentally different beast. BI tools run a predictable set of joins, but an LLM agent is a monkey with a grenade - it can easily hallucinate a 7-way CROSS JOIN or a recursive CTE without a base case. Your read replica will hit OOM or get pinned by I/O in seconds. The only sane path is offloading specific data slices to Snowflake, ClickHouse, or a local DuckDB where you can enforce hard per-user memory and timeout quotas, completely isolating AI noise from your transactional cluster
I'm curious how your analyzers handle those tricky MIPS edge cases like split address loads via lui and addiu where you have interleaved instructions in between. If Ghidra fails to collapse those into a single reference, does your math check just flag it as an error and bail, or does it actually try to reconstruct the instruction chain itself?
Regarding the bots: since you're building a privacy-first product, you should look into a Proof-of-Work captcha (like Hashcash or mCaptcha). Just have the user's browser mine hashes for a couple of seconds before issuing the trial token. A normal human won't even notice it, but it'll burn so many CPU cycles for bot farms that abusing your API becomes economically unviable
Mad respect. I tried extracting a clean .o file out of a statically linked ELF once, and it's an absolute nightmare. How are you handling switch tables and indirect jumps? Without dynamic analysis, it's sometimes physically impossible to figure out what a register is actually pointing to
Thanks for this thread. I'm so tired of reading about yet another "AI wrapper for sending emails" raising a $5M seed round that I almost forgot what actual engineering looks like. The whole industry feels like a Gold Rush right now where 99% of people aren't even trying to mine gold, they're just reselling each other the exact same shovels (LLM APIs) with different logos slapped on them. It's incredibly refreshing to read about people writing Rust parsers or soldering hardware synths
It's also a reminder that there's often a gap between what technology companies market to the masses and what the people behind those technologies actually endorse for their families
It's interesting that the Swedish government isn't completely rejecting digital tools but rather seeking a balanced approach, introducing them when students are ready for them
People are just playing around with parallel agents because it looks cool on Twitter. In real prod, 90% of your time isn't spent typing lines of code, it's spent trying to figure out implicit business requirements and debugging undocumented legacy spaghetti. Agents sitting in isolated worktrees are completely useless here - they'll just rapidly and in parallel write code that completely fails to solve the actual business problem
Spreading agents across git worktrees sounds awesome right up until the merge step. Sure, they're isolated on the filesystem, but when five parallel Claudes rewrite the exact same base class or interface for their own local needs, you're gonna end up with a merge conflict no neural net could ever untangle. All that saved time will just get burned manually rebasing this parallel chaos. Props for a cool pet project, but conceptually this is an architectural dead end
What you're describing (hidden key exchanges with Trust-On-First-Use) is exactly what Signal and WhatsApp already do - they just hid all the math under the hood and tied it to your phone number. A pure Web of Trust where normal people have to manually weigh probabilities is never going to take off. The average user will blindly click "Accept Risk and Continue" on literally any certificate warning just to get back to looking at pictures of their grandkids
Hashing every X seconds is just a Merkle tree, the tech for that has been around forever. But cryptography only protects the container, not the semantic meaning inside it. If verifying a video requires spinning up this massive crypto infrastructure that can just be trivially bypassed with a hardware camera spoofer anyway, that defense is completely worthless for the mass market. Scammers would bypass it in their sleep.
Just go look on HuggingFace. It's packed with uncensored models from the Dolphin Llama 3 70B family that will happily write you a recipe for napalm while swearing like a sailor. Meta's guardrails lasted exactly one week before the community figured out weight abliteration - a method that surgically removes the refusal vectors from the weights without even needing a fine-tune
That only proves the scammer isn't using an OpenAI or Anthropic API. Spinning up Llama 3 70B Uncensored on a rented instance and hooking it up to an unfiltered voice engine is literally a two-hour job. Local weights couldn't care less about morals or safety guardrails
Sure they could, but then you trim the video by 2 seconds, tweak the colors, or just send it over WhatsApp, which recompresses the file with its own encoder. The hash breaks instantly. Cryptography protects bits, but video is about visual meaning. The slightest pixel modification kills the hardware signature. Plus, it does absolutely nothing to fix the "analog hole" problem - a scammer can just point that cryptographically signed iphone camera at a high-quality deepfake playing on a monitor
Picture this: your grandma calls you in a panic, and you tell her, "Drop me your public PGP key so I can verify the signature".. PGP is dead outside of niche geek circles exactly because key management is basically an unsolvable problem for the average person
Those incidents with curl and sqlite were caused by a mob of script kiddies dumping source code into ChatGPT and spamming bug bounties for quick cash. This tool is built by actual Google devs who are active on the kernel mailing list. They know perfectly well that if they start spamming LKML with hallucinations, they'll get patch acceptance blocked for their entire corporation
You don't even need an LLM for this stuff. GCC has the __cleanup__ attribute, and kernel static analyzers like Smatch have been catching missing unlocks for a decade now. People just ignore linter warnings when submitting patches, so the language itself isn't really the issue. The LLM is basically just acting as a talking linter that can explain the error in plain English
Written in Rust, tests a C kernel, using the Google Gemini API... classic 2026. I'd bet 90% of the actual useful work this agent does is just dumb pattern-matching for typical vulnerabilities (use-after-free, uninitialized vars) that the model memorized straight out of the CVE database
That's actually super smart. If you're just piggybacking on native debugger pauses and Mojo IPC instead of hardcoding custom locks into V8, the patchset is genuinely microscopic. I take back the meat grinder comment - with a diff that clean, automated rebases across major Chrome releases are actually totally manageable. Hats off!
I get the readability point, but maybe look into embedded DSLs like they do in Lisp or Elixir? You get your custom, readable syntax inside a macro, but it compiles down to native code so you still get to keep all the tooling and perks of the host ecosystem