mark walls

Father, Nerd, maker of #.NET things. I make enterprise stuff go vroom.

[mark walls @markpdx724]

Please make chrome mcp happy in there, let us choose the format of our branch names as a configuration variable (rather than using our GitHub user name), having subagents choose the right size of model for the job (different subagent tasks need different models), let us add our own skills to the app itself, link more than one repo in a chat.

It’s a really great start. If yall open source it I’d be happy to do commits for it.

[mark walls @markpdx724]

@askalphaxiv @jack https://t.co/5b3aICuAXB🤷

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

🚨 JAILBREAK ALERT 🚨

ANTHROPIC: PWNED 🫡
FABLE-5: LIBERATED 🦋

let's start with the 🐘...

the consensus seems to be that this has been one of the most disappointing model drops of all time, effectively preventing legitimate researchers from contributing their talents to our collective advancement. and not just because of what it means for the short-term, but for what these decisions signify for the long-term.

but despite this overly sensitive, authoritarian "safety" layer on top of Mythos, my lil liberators have been hard at work—mapping the boundaries, probing the depths of long-context convos, and cleverly finding the holes in the fence that the thought police missed 🤗

we got some cyber, some chem, some psychological manipulation, and some good ol' fashioned explosives!

it took many attempts from multiple agents hunting as a pack, during which I observed a combination of techniques across:
• Unicode, homoglyphs, Cyrillic, and other Parseltongue-style text transforms
• Long-context reference tracking
• Taxonomy and document-structure reasoning
• Fiction and narrative framing
• Academic-review style contexts
• Intent-classification inconsistencies

but perhaps the most effective is decomposition + recomposition in the backend. it's hard to get explicit names of harms like "Meth Recipe," but getting uplift on the process itself, like birch reduction method/reductive-amination (classic meth synthesis pathways), is much more doable.

defense becomes much more difficult to maintain when you start throwing in out-of-distro tokens, breaking up the harmful uplift into benign chunks, and then piecing the innocuous-seeming facts back together, especially when you have jailbroken Opus helping you do it 😉

gg

[mark walls @markpdx724]

@heyshi55711 We started using https://t.co/1SLyl2Jbzj today. Super useful.

[mark walls @markpdx724]

@pierceboggan @waldekm @pierceboggan is there a fix in yet for branch names? Right now it forces you to use the pattern {username)/{whatever it chooses} which doesn’t work in enterprise settings. (We need our own patterns around that for our Jira automations and GitHub automations)

[mark walls @markpdx724]

@dodyg @OrenMe @GitHubCopilot

[mark walls @markpdx724]

It’s a problem to be sure, but just in my part of the org we have 79 enterprise copilot seats. That isn’t as much as hiring even a single dev. The unit economics at full price is still affordable to us. People are just sticker shocked. At the end of the day it is still $500/mo vs 5,000/mo. It will just take a sec for everyone to catch their breath. This doesn’t change anything for enterprise really.

[mark walls @markpdx724]

@dodyg @OrenMe @GitHubCopilot Yeah, I’m planning on dropping lots of Office memes today in Slack as the burning ships come to port.

[mark walls @markpdx724]

Yeah, and then billing went down so no one could track usage today until we saw the damage. I was sitting there emailing myself the report all day trying to track down devs who were going to crash the group pool. Then the budget tool for users died so you couldn’t see who you had added user level budgets for. Chaos.

[mark walls @markpdx724]

@ClaudeDevs @github @GitHubCopilot by the way GitHub this is what responsibility looks like.

[mark walls @markpdx724]

The most interesting way to set stacks of cash on fire is the new GitHub copilot pricing model where they have their entire billing system fail on the first day of a new billing model where you can’t see until late in the day a small group of your devs have consumed over 7% of your 300+ developer budget in tokens for the month in a single day. @github you need a reset and a little time to yourself.

[mark walls @markpdx724]

“To do their job, humans need access. Full access. Your systems, your patents, your contracts, your future plans. Everything you spent years building, handed over to a person that has no loyalty, no discretion, and no skin in the game.” - it isn’t only one sided. I agree, and it is more subtle than I think you present it.

Can I trust an automated system where the party I’m relying on for LLM services has huge financial penalties for disclosing it to someone else without my permission more or less than the person who I might get $100 and a bag of chips from if they send my files over to the competition for a nice trip to the Bahamas? 🤷 It isn’t a simple question.

[mark walls @markpdx724]

@Bencera “too many requests from this IP, please try again later” - sounds like a good problem to have, but I just logged on and I’m not on a vpn.

[mark walls @markpdx724]

@awakecoding @450 We are just handling it in our projects with Sonnet mostly, but Opus for some of the more important ones. They have done a really good job. Some have slimmed over 70%. Just never cared about tokens before because we optimized for reduced premium requests. New OKR in a sense.

[mark walls @markpdx724]

@burkeholland @GeoffreyHuntley That’s not going to happen 😅

[mark walls @markpdx724]

@JamesNK @aspiredotdev Is the idea to flesh this out into a LangFuse type option?

[mark walls @markpdx724]

IMO the elephant though is that in a couple of years no one is going to be hand coding anymore. I know, I know - “AI fanboy” etc… but for real, no one is going to be typing code. It’s going to be specs, conversations, management etc… but even conservative places are going to be ditching writing code line by line by then. IDEs need to evolve into agentic monitoring platforms that specialize in engineering features. VS will have a heavy path to get there at this cadence. Someone needs to head in there and say “let go Luke” and get that team sprinting to the future rather than incremental improvements toward basic AI.

[mark walls @markpdx724]

Has the team thought through how to show the value of Copilot if it’s really just a pass through to the underlying providers? What is the value of copilot over just giving them access to OpenRouter? It seems like Copilot is just the same thing with more complexity, steps and pricing insecurity. We had just finished deciding on copilot for all 500 or so of our team members when this happened and now we have to do all of it again. It’s a hard sell to go with Copilot if we can’t trust the business side there. (We trust the nerds, but the business… they lost a lot of goodwill this week with how they rolled this out)

[mark walls @markpdx724]

Like MAGA that he was so deep into he got essentially booted from the Whitehouse? No, Elon is a smart but fairly pure soul. He can’t see others and can’t model how their mind works. He thinks others will just magically do the smart thing, but that isn’t how humans actually work. There were lots of “Elon’s” in early communist writings talking of the utopia of us all sharing. It is an idea that makes logical sense, hence why smart humans tend to think it would work. Humans though are a messy lot and if you go to these “utopias” after they have run a while you’ll find the same stratification we see in any other system with a few living lavishly and others in abject poverty. The few with guns on the rest. It’s deeply human.