Old school Machine learning =good.
Special purpose trained machine learning =ok, getting better all the time.
General purpose AI(LLMs) are trained on garbage, given garbage instructions and therefore regurgitate garbage to folks who donāt appropriately weigh the garbage factor. Thatās extremely destructive.
Aside that, the AI buzzword should be a red flag for over hype to outright fraud.
I started my career fixing windows for workgroups 3.O! i have seen no cell phones to mini pcs everyone has (some have 3 or more). I feel like a.I. Is not going to be used for good, I just read how hackers and scammers are using it to help them. I just read Michigan is getting an a.I. Data center for great lakes cooling! I do not like it much at this time. Too many bad things and not enough good!
Bad:
- Iām concerned about the absurd amount of energy and resources it requires.
- As a software developer, while my job is not in jeopardy, itās going to make it harder for junior developers to get a job when AI can perform many of those tasks.
- It makes stuff up.
- It is biased.
- Itās nice to you so you like it and use it more.
- Big concerns over where the data to train it is coming from and people not getting compensated. It sure seems like stealing.
- Itās getting too big, too fast and Iām pretty sure at some point thereās going to be a collapse.
Good:
Iāve been using Claude to write this wireless pendant Iāve been working on. While it has itās flaws, Iāve been blown away at how good it is. It has saved me a ton of time. Iāve used other ones to create and modify images. You have to be careful, but it can be very useful.
Yep, notwithstanding it has itās purposes (see @jeyeager Jasonās post) if you donāt know the answer, donāt ask AI, if you are looking for an original piece of art you are looking in the wrong place.
The āIā gives everyone exactly the wrong impression, itās not āintelligentā, itās aggregating information.
Silly but real example - A friend wanted to know if her car had one or two reversing lights and stupidly took the google AI result as correct. I posed the same question differently and received the opposite answer. How dangerous can that be?
Seems to me that it emulates Average Intelligence just fine.
Short version: Itās wasteful garbage output, trained on theft, with itās prime usecase being the devaluing of skills, siphoning even more money into the hands of the super rich assholes undermining everything.
Funny thing: experienced people tend to overestimate the time savings due to LLMs, quite drastically even.
I think itās covered above pretty succinctly, but for me itās kinda two key things. The ridiculous environmental footprint and the fact that it does not and maybe cannot sufficiently convey any level of certainty in its results.
From what Iāve seen, great for things that are more āon railsā and flexible output, like writing messages, chat bots, etc. I guess thatās where the coding comes in, as well.
However, itās horrible at hard-fact scenarios in the face of any level of complexity/opaqueness in the topic. Most of the attempts Iāve seen people try to use it for with respect to electronics hardware range from just kinda crappy to outright incorrect/dangerous.
Maybe itāll get there, maybe it wonāt. At the moment I think itās a cute toy with some real world applications, but given human nature and our willingness to seemingly trust it implicitly then I honestly think it could be the herald of the downfall of our civilisation. Cāest la vie, I guess.
On one hand, Cursor (and using Claude inside of it) has transformed my life. Iāve solved soooo many problems at work with python scripts, and Iāve made my own apps and tools with it, so itās more than a time saver for me but itās actually enabled me to follow some coding dreams! After mastering Cursor (everyone should try it cursor.com), Iām pretty much free to explore any possibilities I can imagineā¦
Itās also incredible for non-native English speakers, I have employees who donāt have confidence in their English (despite being amazing at it lol) so they can run something through Claude and feel better about communication before sending things out to a chat/github update.
I do think as a society we need to automate and eliminate mundane tasks. Humans werenāt meant to act like robots on an assembly line all day, every dayā¦ that type of thing is soul crushing. So I say bring that shit on, but also weād need something like UBI to support people. Which again to me is ideal, not everyone is āmeantā to work a 9-5 or any kind of rigid schedule, and I think when we stop forcing people to do so (and successfully help them with UBI etc), we will have a happier population.
I also saw that NVidia is launching an H100 GPU into orbit to test out orbiting data centers. Free, unlimited electricity from the sun, and no fresh water used in cooling. And it propels us further into space! Couple that with the moon bases and we could be fairly close to mining the moon and manufacturing the GPUs etc on the moon. Then we can have our cake (AI) and eat it too (no resource depletion, advancing space tech). For a fascinating and thrilling book about a realistic path to space mining and manufacturing, check out Delta V by Daniel Suarez (audiobook is extra good). I also think the energy demand for terrestrial data centers is pushing a resurgence in nuclear energy, which Iām a tremendous fan of. So in summary, demand for AI is already propelling us towards futuristic technological growth, and I think that kicks ass.
On the other hand, I absolutely do not think LLMs should be freely available for everyone and their mom to use with every single Google search, shitty AI filters/image generators etc. Meta is building a new datacenter a couple miles from my work and itās fucking enormous, and itās only going to employ like 100 people, and will be sucking up hella fresh water for cooling. And Iām sure they wonāt be using all that compute power to cure cancer, but to instead make shitty memes and develop algorithms to sell to the DOD etc. but Iād love to be wrong 
I sure wish we didnāt let them get away with calling LLMs āAIā. Weāre going to feel silly when the real stuff comes out and we have to come up with an even better term for it.
Not trying to pick on you, Jono. But it is pretty good at facts that are common knowledge. It goes wrong the same way non experts do, spouting myths and guesses as facts when we either donāt know the answer, or the correct answer is swamped with misinformation. Basic truths are generally right. Although admittedly, I donāt ask it a lot of questions that I already know the answer to.
I do appreciate that newer search responses include sources and I can follow them to get the real info and see if I trust the source. LLM is better at understanding my questions than the regular search engine sometimes. The top responses will be related, but not helpful and the LLM response gets the question right, and the sources are what Iām looking for.
In programming, it seems good at doing basic mundane things, but not as good at complex, difficult, or corner cases. But Iāve always said, coding is a tool. I was never hired for my ability to code. I was hired for my ability to make robots work with coding. Just like you donāt hire a woodworker because they know how to use a table saw, you hire them to build a bookcase. The subject matter they are an expert in is making furniture. Using tools is just how they are effective at it. Future coders need to focus on being an expert at solving problems with LLMs, not the esoteric workings of a compiler (unless you are working on a compiler).
It is a stretch to think that this level of LLMs is on an exponential trajectory towards super intelligence. These versions are an average of our knowledge. They are encouraged to average that intelligence. It will be a big leap to make it an order of magnitude smarter than us, in the realms we are already good at.
It will happen. But we need a differenr approach, IMHO. Humans donāt push the boundaries by listening to lectures from experts. They push the boundaries by experiements and creative thought. LLMs arenāt capable of that.
I disagree here partially.
Knowing when to use an LLM is going to be important. Knowing when you should stop using one and look into specialized sol tools as well.
Knowing when not to use one despite it doing just fine, so youāll be able to fix it once it inevitable fucks up, that is going to be a problem.
Work has us ātesting AI outā and using it as often as we can. The comment above about AI being a replacement for a Jr dev is pretty accurate for me. Could I write a sed command to replace this variable in all my codeā¦ yes. can I also just ask claude to do it? yes. Does claudeās code always work. no.
Most of my coding is python, terraform, and helm charts. Claude has done well for making quick changes and updates to existing code for me. I have not used it for any extreme logic or odd asks. Mostly keep using it for minor changes and updates. āClaude, exclude āthis tagā from all my boto3 results in this scriptā.
For me, AI is a lot like CNC. Iām using it for repetitive stuff, but Iām still handling the initial design.
Itās also been really good at condensing tasks for me. Work recently sent out a new handbook I had to sign. I took the old PDF and the new PDF and had Claude summarize the differences. Iāve used it similarly to compare electronic devices. Could I scour the web looking for spec sheets and doing the work? yes. I have used the google AI driven answer as a start of my personal research on a few questions. It, again, seems to do ok at summarizing websites and at least somewhat pointing me in the right direction.
Iāve also seen the bad side of AI where a long running discussion starts to cause hallucination on the AI side. Once that happens, itās best to start over with a new topic.
Iāve tried running my own local LLM. I donāt have a GPU on the system it was running on, so it was really slow. It was at least consistent. I was trying to write a phone app that I could take a picture of a cookbook page and have it import it into my personal Mealie recipes database. I donāt need that to be very fast, just accurate. I ran into issues and havenāt taken the time to try to figure them out.
My wife uses AI a lot to help her write emails to parents. Sheāll give the AI a bunch of bullet points for what the email needs to talk about and then has it fill in the word-smithing. Iām kinda happy sheās doing it as I used to do all her word-smithing for her.
Out of my league here and donāt really know the diff between LLM and AIā¦ and Iām definitely an age-related corner case. But I find Gemini/Grok really handy to ājog my memoryā (which needs a lot of ājoggingā nowadaysā¦) and help walk me through procedures that I used to do routinely and oftenā¦ but now canāt remember all the steps or syntax while still knowing basically what I want to do. I can remember a few bits and piecesā¦ but need the reminder to check the bits Iāve forgotten.
I encourage you all to take the time to read this short story called āThe discrete charm of the Turing Machineā
Considering it was written in 2017 I think itās still very very relevant (and disturbingly prescient)
Not feeling picked on at all.
This is why I said 'in the face of any level of complexity. Although even on here Iāve still seen it go wrong about using graphite as a lubricant in a way that I could find zero reference to support the LLM output.
And therein lies my problem. You may be using it ācorrectlyā. What I seem to mostly see is people using it incorrectly. We already put way too much stock in stuff that we shouldnāt, like the accuracy of reporting/media, the opinions of celebrities, advertising etc. Adding an over-confident LLM to the pile and combining it with people who see it be right just often enough to go āwell, surely itās right about this, tooā, is what concerns me. Itās the āsunlight and bleachā approach.
That tracks with what Iāve experienced, professionally, except that I think my area of interest is a worse fit for it in general because it doesnāt distill well down to text, which seems to be the crux. From what Iāve seen, itās weirdly bad at interpreting accurate information of graphs if there is any level of inference required.
This is an example of where the discussion crosses into hype territory, and where lack of knowledge is particularly dangerous. I happen to know a bit about this topic, (my āreal jobā is building, testing, and flying interplanetary spacecraft).
It isnāt free to launch a spacecraft, and although solar energy is present, the vehicle needs to be built, launched, and operated. Not free, and not cheap even with the best of the emerging low-cost reusable launch vehicles and associated extreme low-cost spacecraft.
Cooling isnāt free in space, you have to radiate the waste heat as you have neither convection nor conduction available in the vacuum of space. That means that per unit of power dissipated itās actually considerably more expensive and more difficult in space.
And then, worst of all, the numbers being bantied about for compute power requirements in the coming years adds up to more than the entire available electrical power generation on the surface of the planet today. That isnāt viable either on this planet or in orbit of it.
(This mostly ignores the very real possibility of the massive constellations of spacecraft rendering near earth orbits unusable from cascading debris generation, which is a still worse problem.)
Youāre in the realm of impossibility here. So. much like āFull self drivingā for cars, this is a complete bullshit hype story that plays on the gullibility of an uneducated populace as a part of inflating stock prices using plausible sounding but bogus predictions.
Well you just succinctly stated Elonās business model for just about every company that he is associated with (Tesla, Space X, Boring Company, etc.)
SpaceX is legit. Thatās because the adults keep the head at appropriate length from operations.
But, Mars is to SpaceX as FSD is to Tesla.
One of my side things is writing.
Among authors, Iāve heard many times that āevery story has already been told.ā AI or earher LLMs are or at least could be an excellent tool to remix other story elements onto a new retelling of existing plots, but with no real possibility of coming up with an actual new story.
While every story is new in detailed composition, as should be an LLM creation, the basic plots usually already exist.
Take that to software creation. That is my work life, or at least a large part of it. When I started programming, we still spent lots of time dealing with some of the basic building blocks: the ones that OOP was supposed to relieve us of, but didnāt. Weāre starting to not have to worry about those anymore, but pay for it with inefficiencies. A LLM can probably recover some of those inefficiencies, if properly trained. That can free up significant computing resources, used properly.
Will they be used properly? Sure. Will thry be used improperly? Absolutely. On balance, the trick is going to be weeding through the people who donāt know the difference.
True AI has yet to come about, and itās going to be a whole new can of worms.