Lack of versioning in workspaces. Multiple times I've been "vibe coding" my way happily through adding feature after feature until ... the chatbot goes off the rails and breaks stuff. No amount of additional prompting will fix the issue no matter how many time Claude/CoPilot says "I see the problem. Let me fix that..." and it only gets worse.
Claude provides a * SINGLE STEP * of undo. Gee, thanks.
So, now I'm forced to snapshot code myself. I just wish that was built in.
The death of entry-level math and CS jobs is greatly exaggerated, I believe.
First, costs and ROI of AI are on a collision course and companies are finding they can't simply lay off smart people and replace them with "vibe-working" employees. Secondly, not only will more junior workers adapt to new hybrid human/AI workflows better than more senior workers, they're likelier to apply them to parts of the job that are low-value, routine. Third, the next frontier for agents and models is within very narrow domains where more junior AI-savvy workers are in a great position to collect, capture and refine the data models from subject matter experts. Who better to ask "What about X ...?" and "How do you know Y...?" from experts?
1. AI/ML is not an oracle. It is neither wise nor omniscient. Never rely on AI for advice (I tell them the pizza glue story). Never rely on AI for opinion (because if you ask for a different opinion it will deliver it, happily).
2. AI/ML is not a source. It can helpfully summarize things and tell you things that may be true but then you need to seek out the sources both to smoke out the errors and hallucinations but also to be exposed to gritty details that get sandpapered away by the "smoothing" of the ML models. AI output tends to be like those posts that show the average face for every country. The resulting faces are interesting and pleasant but should never be confused with real people.
3. Generative AI (text, images, videos, code) is an amazing and fun way to turn an idea into something concrete. But ... beware the pitfalls. A) Beware the delight of novelty. Even if the image/text is kind of coherent and novel, scrutinize the details. Did the result express your original intent?
B) Sweat the details. If the result doesn't show/say 100% of what you wanted then it isn't 100% yours. If you stopped there, you would be letting the AI/ML dictate the limits and "flavor" of your creation. It would be better to use the generated output as inspiration or a guide to what you ultimately create.
I like your argument, not because it is a good analogy for AI but because it is a good contrast. Copyright isn't a guarantee or magic force field blocking fair competition. It is a permeable buffer against lazy knockoffs and time-boxed so that buffer doesn't choke all future creativity.
People on this thread need to focus on what "derivative" and "fair use" mean and understand both are measured on a somewhat fuzzy spectrum, subject to interpretation.
In a perfectly fair world AIs/MLs could vacuum up all human knowledge, fair and square. (In an ideal world, they would do that adhering to polite opt-in/opt-out agreements with copyright holders. We can dream). Input isn't theft.
On output, two magic genies would stand at the gate, the Derivative Genie and Fair Use Genie and review anything spat out by the AI/ML. If it crossed agreed upon thresholds the Genies would bar the gates and issue a stern warning to prompt again (or maybe the AL/ML would auto-adjust the prompt and try again).
So, if your prompt asked for a 300-word poem about thrash metal mosh pit dancing and it spat out a poem where 85% of it match one of the handful of available mosh pit poems in its database, the Derivative Demon would block the output and raise an alarm.
On the other hand, if you asked for a line by line analysis of a famous mosh pit dancing poem (by name) or maybe asked for a satirical spoof of said poem, the Fair Use Demon would overrule the Derivative Demon and give the output a pass.
That's as fair as this could get, especially if you add one more thing: An Appeals Court (maybe corporate, maybe 3rd party, maybe state run) with a Settlement Pool. If a copyright holder could prove the Genies let pass something they shouldn't, the AL/ML would fix that. If real damage is done, the creator would get a settlement from the pool.
The point is that the Input Genie is out of the bottle. Creators just look foolish trying to squeeze it back in. Better, they should focus on making the output Genies and the Appeals process as effective and fair as possible for everyone.
Does he WANT worldwide eternal Jihad against the US/West?
If there's a better use for "Some people just want to watch the world burn" in the history of the world, I can't think of one. He's always been a toddler brain clueless to consequences and, in this case, he likes the idea that HE would get to be the Big Boy to push the Big Red Button. Fun toy.
- I'd like to see this as a hosted app versus something that has to be "installed" in a chatbot.
- It needs a text search feature for both the outline and full text. That would allow searching for text containing "government", highlight instances and seeing their context. And same for searching the outline for "government" and seeing supporting text.
This could be an equally useful paradigm for fiction and for source code. For fiction, it would be really useful if this could be trained to identify character introductions and locations and their mentions. Imagine how convenient it would be for the outline to mention a plot point about "Mary Sullivan" where the paragraph in chapter 22 only says "his mother" when talking about "John Sullivan."
I've thought about a hosted app and it's something I'd like to do eventually. The API costs and building a reliable service make it a bigger project though, so it's not happening soon.
Text search is a good idea. It makes reading more active because you pick a concept, see where it shows up and how it fits the structure. The fiction use case you mentioned is similar in a way, finding where a character or concept appears even when the text doesn't use the exact same words. I'll look into it.
At minimum, you have to provide a snapshot/example of what you're building.
I'm most intrigued by your comment pile digestor. I've often wanted something like that while cruising the Internet but I'm not sure how'd you monetize it. In the commercial world, software for analyzing customer-submitted questions and comments has been a valuable service for a long time, even before AI/ML engines came along. So, that's useful but it's a hard problem, very domain-specific.
Anyway, as others have said the main things you have to do to attract interest/business are: 1) Explain specifically what pain you are solving for what specific user/customer, 2) Offer specifics (examples, features) of how you solve that pain better than anyone else.
I agree with you this would be valuable, mostly for persuading the general public.
The skepticism you're getting amounts to "biased people/government won't care no matter what" and "fancy tech makes the evidence look more doctored, not less." They're not wrong but there are some disputes this could settle. Some people have argued that the moment where an ICE agent appears to withdraw Pretti's gun might just be him retrieving his own fumbled gun. Maybe a synthesized view could clear that up.
Personally, I think it will come down to the moment where one goon said "He's got a gun" and another one took that as license to shoot him three times. The other 6 or so gunshots were a kneejerk pile-on to that cowardly act.
So much for the 2nd Amendment, right?
Right, the analogy would be "What does being a builder mean if carpenters are swinging the hammers?" We'll have to be software "Foremen" who develop plans and know enough algorithms, data structures and architecture to tell the AI how to frame the house and what kind of pipe to lay even if we're not doing that ourselves.
Claude provides a * SINGLE STEP * of undo. Gee, thanks.
So, now I'm forced to snapshot code myself. I just wish that was built in.
reply