[HN Gopher] Things I've learned in my 7 years implementing AI
___________________________________________________________________
Things I've learned in my 7 years implementing AI
Author : jampa
Score : 113 points
Date : 2025-10-15 18:27 UTC (4 hours ago)
(HTM) web link (www.jampa.dev)
(TXT) w3m dump (www.jampa.dev)
| nh23423fefe wrote:
| nit: space bar doesn't scroll
|
| this list isnt learnings.
| sodapopcan wrote:
| It doesn't. It's a jump scare, and a stupid one at that.
| jampa wrote:
| I'm not sure what you mean. This is the vanilla Substack
| layout.
|
| EDIT: Removed the video because a bug in Substack causes the
| space bar to play the video instead of scrolling down. Sorry
| for the unintentional jumpscare.
| CamperBob2 wrote:
| Seems to make noise (I closed the page ASAP so not sure
| what it actually did.)
| stronglikedan wrote:
| It means scrolling with the space bar is janky and broken,
| instead of smooth and working properly.
| jampa wrote:
| Removed the video because a bug in Substack causes the
| space bar to play the video instead of scrolling down.
| Sorry for the unintentional jumpscare.
| sodapopcan wrote:
| Oh I didn't even see my video. I honestly thought it was a
| sick joke, lol. My speakers were turned way up.
| SeanAnderson wrote:
| holy shit lol. I should've trusted you but I hit spacebar and
| was NOT expecting that hahaha.
| jay_kyburz wrote:
| Oh, I wanted to see what you guys are talking about, but I
| just get a scroll. I wonder if they changed it already.
| wppick wrote:
| The one issue is the accuracy of these AI models, which is that
| you can't -really- trust them to do a task fully, so that makes
| it hard to fully automate things with them. But the other is
| cost. Anyone using these models to do something at scale is
| paying maybe 100X over would it would cost in compute to run
| deterministic code to do the same thing. So in cases where you
| can write deterministic code to do something, or build a UI for a
| user to do it themselves, that still seems to be the best way.
| Once AI gets to the point where you can fully trust some model,
| then we've probably already hit AGI and at that point we're
| probably all in pods with a cable in our brainstems, so who
| cares...
| dayvid wrote:
| The article kind of addresses that in identifying what are the
| best type of problems AI can solve
| frankc wrote:
| The thing is that I don't use AI to replace things I can do
| deterministically with code. I use it to replace things I
| cannot do deterministically with code - often something I would
| have a person do. People are also fallible and can't be
| completely trusted to do the thing exactly right. I think it
| works very well for things that have a human in the loop, like
| codeing agents where someone needs to review changes. For
| instance, I put an agent in a tool for generating aws access
| policies from english descriptions or answering questions about
| current access (where they agent has access to tools to see
| current users, buckets policies etc). I don't trust the agent
| to do it exactly right so it just proposes the policies and I
| have to accept or modify them before they are applied, but its
| still better than writing them myself. And it's better than
| having a web interface do it because that is lacking context.
|
| I think it's a good example of the kind of internal tools the
| article is talking about. I would not have spent the time to
| build this without claude making it much faster to build stand-
| alone projects and I would not have the agent to do the english
| -> policy output with LLMs.
| enraged_camel wrote:
| >> The thing is that I don't use AI to replace things I can
| do deterministically with code. I use it to replace things I
| cannot do deterministically with code - often something I
| would have a person do.
|
| Nailed it. And the thing is, you can (and should) still have
| deterministic guard rails around AI! Things like
| normalization, data mapping, validations etc. protect against
| hallucinations and help ensure AI's output follows your
| business rules.
| Terr_ wrote:
| > Things like normalization, data mapping, validations etc.
| protect against hallucinations
|
| And further downstream: Audit trails, human sign-offs,
| operations which are reversible or have another workflow
| for making compensating actions to fix it up.
| asdff wrote:
| Or, you could make a tool that can generate this stuff
| deterministically every time the exact same way. At least
| with that situation you can audit the tool and see if it is
| correct or not. You still leave the point of failure on the
| user in your situation, even higher because they could get
| complacent with the llm output and assume it is correct or
| mistakenly think it is correct.
|
| In my mind you are trading potentially a function that always
| evaluates the same for a given f(x) for one that might not
| evaluate the same and requires oversight.
| basisword wrote:
| >> What I do see is a boom in internal tools.
|
| This has been my main use case for AI. I have lots of ideas for
| little tools to take some of the drudgery out of regular work
| tasks. I'm a developer and could build them but I don't have the
| time. However, they're simple enough that I can throw them
| together in a basic script form really quickly with Cursor.
| Recently I built a tool to analyse some files, pull out data, and
| give me it in the format I needed. A relatively simple python
| script. Then I used Cursor to put it together with a simple file
| input UI in an electron app so I could easily share it with
| colleagues. Like I say, I've been developer for a long time but
| never written python or packaged an electron app and this made it
| so easy. The whole thing took less than 20mins and it was quick
| enough that I could do it as part of the task I was doing anyway
| rather than additional work I needed to find time to do.
| isoprophlex wrote:
| > Creating AI models is hard, but working with them is simple
|
| I'm not disagreeing with the overall post, but from closely
| observing end users of LLM-backed products for a while now, I
| think this needs nuance.
|
| The average joe, be it a developer, random business type, a
| school teacher or your mum, is very bad at telling an llm what it
| should do.
|
| - In general people are bad at expressing their thoughts and
| desires clearly. Frontier LLMs are still mostly sycophantic, so
| in absence of clear instructions they will make up things. People
| are prone to treating the LLM as a mind reader, without
| critically assessing if their prompts are self-contained and
| sufficiently detailed.
|
| - People are pretty bad at estimating what kind of data an LLM
| understands well. In general data literacy, and basic data
| manipulation skills, are beneficial when the use case requires
| operating on data besides natural language prompts. This is not a
| given across user bases.
|
| - Very few people have a sensible working model of what goes on
| in an autoregressive black box, so they have no intuition on
| managing context
|
| User education still has a long way to go, and IMO is a big
| determining factor in people getting any use at all from the
| shiny new AI stuff that gets slathered onto every single software
| product these days
| dayvid wrote:
| Prompt engineering is a transitory phase. Embedding it into
| existing tools so the 80-90% of regular prompt patterns can be
| worked into the UI (or contextual UI designed around how a user
| uses the product) is the next step
| isoprophlex wrote:
| Yeah if you mean that LLMs are used with UIs on top that
| allow "too much magic", I agree.
|
| Free form chat is pretty terrible. People just want the thing
| to (smartly) take actions. One or two buttons that _do the
| thing_ , no prompting involved, is much less complicated.
| gubicle wrote:
| The whole point is the prompt (+ a static set of
| (system)prompts). If your whole function as a human is
| clicking one of a set of buttons to trigger an AI action,
| then you are automate-able in a few lines of code (and the
| AI is better than you at deciding which button to click
| anyway (supposedly)).
|
| There are like thousands wrappers around LLMs masquerading
| as AI apps for specialized usecases, but the real
| performance of these apps is really only bottlenecked by
| the LLM performance, and their UIs generally only get in
| the way of the direct LLM access/feedback loop.
|
| To work with LLMs effectively you need to understand how to
| craft good prompts, and how to read/debug the responses.
| wppick wrote:
| The higher your executive function the more use you will get
| out of LLMs. This is really the skill you should be testing for
| in interviews now. Not letting a candidate use AI for their
| interview is not a useful evaluation anymore. I want to see how
| you use it. Do you prompt well, how much are you trusting and
| verifying what it outputs
| procaryote wrote:
| Is "executive function" the thing you mean here? You need a
| lot of self control to write good prompts?
| dgfitz wrote:
| Is this a "you're holding it wrong, you idiot" post?
| isoprophlex wrote:
| No no definitely not meant to be condescending; our ui
| paradigms don't align with users that just want to get stuff
| done. See what dayvid writes, too.
|
| and everyone's being sold on this tech being super magic but
| to some questions there is an irreducible complexity that you
| have to deal with, and that still takes effort.
| random3 wrote:
| I like the engineering part at the top but projecting AI
| perspectives blindsided through the lens of LLMs is effectively
| "looking backwards".
|
| So this is nice
|
| > productionizing their proof-of-concept code and turning it into
| something people could actually use.
|
| because it's so easy to glamorize research, while ignoring what
| actually makes ideas products.
|
| This is also the problem. It's a looking back perspective and
| it's so easy to be miss the forest from the trees when you're
| down in the weeds. I'm talking from experience and it's a feeling
| I get when reading the post.
|
| In the grand scheme of things our current "AI" will probably look
| like a weird detour.
|
| Note that a lot of these perspectives are presented (and thought)
| without a timeline in mind. We're actually witnessing timelines
| getting compressed. It's easy to see the effects of one track
| while missing the general trend.
|
| This take is looking at (arguably "over") LLM timeline, while
| missing everything else that is happening.
| only-one1701 wrote:
| Ok, so if projecting through the lens of LLMs is looking
| backwards...what would you have preferred to see?
| cadamsdotcom wrote:
| > What I do see is a boom in internal tools.
|
| It's easy now to get something good enough for use by you,
| friends, colleagues etc.
|
| As it's always been, developing an actual product is _at least_
| one order of magnitude more work. Maybe two.
|
| But both internal tools and full products are made one order of
| magnitude easier by AI. Whole products can be made by tiny teams.
| And that's amazing for the world.
| zwnow wrote:
| > Everything got one order of magnitude easier thanks to AI.
|
| No. Not at all. Many things maybe got easier but a lot of
| things got magnitudes harder. Maintaining bug bounty programs
| for example, or checking the authenticity and validity of
| written content on blogs.
|
| Calling LLMs are a huge win for humanity is incredibly naive
| given we dont know the long term effects these tools are having
| on creativity in online spaces, authenticity of user bases, etc
| etc.
| cadamsdotcom wrote:
| A thing doesn't have to be universally good to be good.
| zwnow wrote:
| Obviously, but id refrain from calling something good
| without knowing the extend of its damage.
| reppap wrote:
| Agreed. We don't know if AI is another asbestos or not,
| yet.
| asdff wrote:
| Whole products could always be made by tiny teams.
| asdev wrote:
| You can't rely on them directly for business value, but you can
| rely on them indirectly is the summary of this post
| aDyslecticCrow wrote:
| This article aligns very well to my frustration to the current
| view of AI in media and discussion.
|
| > AI tools like KNNs are very limited but still valuable today.
|
| I've seen discussions calling even feed-forward CNNs, monte-carlo
| chains, or GANs "antiquated" because transformers and diffusion
| have surpassed their performance on many domains. There is a
| hyper-fixation on large transformers and a sentiment that it
| somehow replaces everything that came before in every domain.
|
| It's a tool that unlocks things we could not do before. But it
| doesn't do everything better. It does plenty of things worse (at-
| least taking power and compute into account). Even if it can do
| algebraic now (as is so proudly proclaimed in the benchmarks),
| wolfram alpha remains and will continue to remain far more suited
| to the task. Even if it can write code; it does NOT replace
| programming languages as I've seen people claim in very recent
| posts on here on HN.
| groby_b wrote:
| Many good points in the article, but I'd caveat that if you judge
| performance only by ELO score, you are not applying the best
| criteria.
|
| SWEBench performance moved from 25% to 70% solved since beginning
| of 2025, and even with the narrowest possible lens from 65% to
| 70% since May. ARC-AGI2 keeps rapidly climing. We have
| experimental models able to (maybe) hold their ground at IMO
| gold. As well as performing at IPhO gold level.
|
| And that leaves out the point that LMArena is a popularity
| contest. Not "was this correct", but "which answer did you like
| better". The thing that brought us glazing. A leveling ELO (or a
| very slowly climbing one) is kind of expected, and is really
| saying nothing about progress in the field.
|
| Still doesn't mean "THE MACHINE GODS ARE COMING", but I would
| expect to see continued if slowing improvement. (I mean, how much
| better can you get at math if you already can be a useful
| assistant to Terence Tao and win IMO gold? And wouldn't we expect
| that progress to slow?)
|
| But more than "how hard of a problem can you solve", I expect
| we'll see a shift to instead looking at missing capabilities.
| E.g. memory - currently, you can tell Claude 500 times to use uv,
| not pip, and it will cheerfully not know that again the 501st
| time. That's much more important than "oh, it now solves slightly
| harder problems most of us don't have". And if you look at arxiv
| papers, a lot are peeking in that direction.
|
| I'd also expect work on efficiency. "can we not make it cost
| about the amount of India's budget to move the industry forward,
| every year" is kind of a nice idea. No, I'm not making the number
| up, or taking OAIs fantasy numbers - 2025 AI industry capex is
| expected to be $375B. We'll likely need that efficiency if we
| want to get significantly better at difficulty level or task
| length, too.
| Nevermark wrote:
| > AI as a product isn't viable: It's either a tool or a feature
|
| This correlates with the natural world. Intelligence isn't a
| direct means of survival for anything. It isn't a requirement for
| physical health.
|
| It is an indirect means, I.e. a tool.
| blobbers wrote:
| "There's a reason we're not seeing a "Startup Boom" AI skeptics
| ask, "If AI is so good, why don't we see a lot of new startups?"
| Ask any founder. Coding isn't even close to the most challenging
| part of creating a startup."
|
| -- uhhh... am I the only one seeing a startup boom??? There are a
| bajillion kids working on AI start ups these days.
| at-fates-hands wrote:
| Nope, you're spot on.
|
| _AI deals continued to dominate venture funding during the
| third quarter. AI companies raised $19 billion in Q3, according
| to Crunchbase data. That figure represents 28% of all venture
| funding._
|
| _The fourth quarter of 2024 has been no less busy for these
| outsized rounds. Elon Musk's xAI raised a behemoth $6 billion
| round, one of seven AI funding rounds over $1 billion in 2024,
| in November. That's just months after OpenAI raised its $6.6
| billion round._
|
| https://techcrunch.com/2024/12/20/heres-the-full-list-of-49-...
| asdff wrote:
| Startups optimize for things that secure initial seed money not
| things that stand up on their own two feet. That is why most
| fail. What VCs invest in and what succeeds is not really
| correlated.
___________________________________________________________________
(page generated 2025-10-15 23:00 UTC)