Yeah, this recent flurry is a batch of colleagues at an overseas office who took an interest in running a procedure I setup about 3 years ago, they’re all creating their own ssh keys and I’m adding their public keys to ~/.ssh/authorized_keys on a certain server - so far so good, but for some reason they’re all renaming their key files away from the default names even though the procedure shows steps to create default named keys (they don’t have other keys…) so, this morning I went and edited the wiki page to make it explicit: “Use the default key names unless you want to do a whole lot more configuration work on your end…” which, of course, none of them do.

But, yesterday, when this query-storm started, they sent me a screenshot with their renamed keys and I wasn’t sure if that should work or not, so I fed the screenshot over to GPT and it instantly answered about how to configure the client to use non-default key names - which I translated to the users as: “just rename your keys to the default names” they did, and they’re happy. Though, this morning one sent me a key both renamed and with her user ID stripped off - I wasn’t sure if that would work or not so we just tried it, and when it didn’t work I had her send me the whole key…

A lot of what I get out of GPT is stuff I already know, or knew, but can’t articulate all the detail about as quickly as GPT can.

even more difficult days could be ahead

The beatings will continue until morale improves.

Yes, do you answer questions for money outside of work? Outside of work if somebody is asking me a question I assume they want my answer and I’ll give them that instead of looking something up, although sometimes I punt with an “I don’t know but I bet Google does…” Inside work I attempt to answer questions as correctly and efficiently as possible - the GPT tools are great at that.

what the kids themselves are saying.

What the kids themselves are saying, or what they’re actually doing?

Self-reported behavior among teenagers is… inaccurate at best.

Well, when the question is: why isn’t my server access working, and the result from gpt gets their server access working… I hope you can trust a result like that?

It’s not that I don’t know, it’s that I’ve already answered their questions, in writing, if they would just read a half page of text and do what it says.

I get questions like: why can’t I access this server, I followed the wiki page (first clue, they didn’t follow the wiki page). That’s not asking for insight, that’s asking for where they failed to follow a set of 5 step directions by doing things like: changing the default filename of their new ssh key to something they invented.

GPT explained, far more patiently than I would have, how indeed to do 4 more steps and rename your ssh key to anything you want, but I did offer the insight: if you just leave the name as the default value, you can skip all of this extra work.

They’re really good at digging for stuff, like: this app is reporting the git hash it was built from - somewhere in the log files - go read that and show me which branch that hash appears on (hash is 8 commits back in some branch…) Yeah, I could do that myself, but why would I if I don’t have to?

Increasingly, people ask me questions, send me screen shots, I copy-paste that into gpt, gpt’s answers are helpful and correct… they have access to the same (free to use) gpt themselves…

Teams will be appropriately sanitized and think you are still at home.

I wouldn’t count on that in our IT environment. The one constant there is: change. Whatever works one week, they’ll screw around with it and make it different the week after that… I WFH and mostly on my own gear because getting things done with their supplied laptop is 50% fighting their ever-shifting “support.”

grab a router and install OpenWRT onto it, and turn it into a wireless bridge. Use the router to connect to the office WiFi, and have a wired connection to your laptop

A setup like that would trip our IT security. I actually killed all the phones on our floor by doing a simple packet route they didn’t recognize, they have the routers set to kill all the PoE and block all data when anything they don’t recognize comes through.

Yeah, that was a decade+ long running cluster of epic proportions… There had to be ulterior motive, like gaining insight into the products in the market or something. Also, the language of the encryption blockade did encourage a lot of products to just use insecure encryption instead of cranking up the “bits” of security to impossible to crack and tangling with the export controls people.

It’s not like the blockade stopped a single copy of anything from moving out of the country, but it did shape the mass marketplace to be less secure.

I miss the days when the SEO bots weren’t winning.

The ruling is flawed, searching the Internet has been an “AI” battle for 20 years using the predecessors of LLMs to sort out “what people really want” vs the websites that are precision honed to receive as many top-ranking search result returns as possible. Then, of course, Google forgot the n’t in “At least don’t be Evil.” and they started pushing promoted (aka paying customers’) results higher in the rankings.

If you simply unplug Gemini, what replaces it? Is Hadoop “too smart” for the ruling? Multiple cross references of content and links and what all else proprietary algorithms the Google goblins cooked up over the last 20 years, at what point is that AI/not AI? If Gemini gets repackaged as “totally not AI tech” - does that make it now legal?

People do need to curb their enthusiasm, on both sides of the AI questions. It’s a tool, it’s not perfect for everything, it is good for some things, better than the best of what came before - for some things.

remind yourself of things you already know (what was the command for X again)

Speak for yourself, they remind me of things I used to know. I have reached a point where I feel like I have forgotten more than most people know.

This isn’t new since ChatGPT and friends dropped. For years before that, Google search results did limited interpretation of natural language requests, not just keyword match frequency. The SEO arms race drove a different kind of AI in search fetching for at least a decade before natural language chatbot tech hit the scene.

I don’t know how much is intentional enshittification to make AI results look better vs how much is simple neglect of the SEO arms race vs maybe it’s genuinely getting harder to deliver good simple search results with LLMs acting as SEO agents?

What I do know is: “AI Mode” delivers more useful information than the old style page link list does these days. The pages linked from the AI Mode results tend to be relevant and useful more than the top page of page links. Hallucinations are way down from where they were 2+ years ago, even better than “top results” misses used to be, IMO. If you’re not getting enough sources in your first AI mode response, ask for more - it delivers.

As was true since the first days of the internet: trust nothing. This is random junk people stick on the web for their own purposes, you have been warned.

US, UK, tale as old as time really - how to stay powerful? Undermine your competition.

Again. They’re on the rinse-lather-repeat cycle for 400 years now.

Google quotes a standard Gemini query at 0.24Wh - and I’ll say if you’re continuously asking normal questions and getting answers at normal speed from Gemini, you might get 100 queries in per hour - so, at that rate, Gemini is consuming 24 watts while in use.

Interestingly, the human brain also consumes about 20 watts, so I’m here wondering if Gemini is cooking its own numbers on the first response.

When you ask it complex questions, it takes longer to respond, but says they might range up to 15Wh per response, so maybe more on the order of 500W while in continuous use for complex queries - like the power of 25 human brains instead of one.

Of course, human watts come from direct digestion of rice and beans and other “solar powered” energy sources, while electricity comes from more environmentally challenging sources.

whether it’s worth the effort to do a thorough review.

If the vibe coder learns how to vibe better…

I’ve been using LLMs for a lot of things since last October, the models have improved pretty dramatically since then, but so have my skills in using them - so it’s hard to tell (and probably unimportant) which factor is more important in the increased quality and efficiency of my code production and reviews over the last year.

Using LLMs to review code (regardless of who/what wrote it) is a more efficient way to improving code quality, security, maintainability, etc. than just reading it all yourself. Certainly don’t go blindly trusting the LLM reviews, but if you haven’t tried them for pull request review, you should…

If you’re going to TOR and discard all tracking cookies, and only use anonymized contact methods (not e-mail accounts linked to your credit cards…) then you’re on the right track.