Csaba Kétszeri

@martinvars Experience on working with junior staff became more valuable. My agents also love to get their job done the easy way when they sense monitoring is not that thorough as it was at the initial period.

The honest secret of running AI agents inside a real company is this: the model is not the bottleneck anymore. The bottleneck is what happens when the agent is wrong. I run agents across several of my companies. They sort emails, manage dashboards, block bots on X, draft replies, summarize calls. The first version is always magical. The tenth version is where you learn the real lessons. The model is rarely the problem. The problem is that nothing in the stack tells you, in production, that the agent quietly drifted. It does not crash. It does not error. It just becomes slowly worse at the job, and three weeks later you realize half of its outputs are subtly wrong. What you actually need is unglamorous: evals you trust, logs you can search, the ability to roll a single agent back to last week, and a human review queue for anything that touches money, legal text or a customer. Most teams skip all four because they are not as fun as a new model. The companies that win with agents will not be the ones with the smartest model. They will be the ones whose engineers treat agents like junior employees with bad memory and worse judgment, and build the supervision around them accordingly. Intelligence is cheap now. Accountability is what will be priced.

@icanvardar Remove all questions answered in a passive-aggressive way that it is a duplicate and the user asking is a moron, then you will get a very different shape. SO was killed by the culture, not by AI.

stackoverflow raised developers developers raised ai ai made developers stop needing stackoverflow and now it’s getting quietly erased by the thing it helped create

@BoringBiz_ That the growth in demand is not necessarily following the growth of supply.

The AI labor replacement theory makes absolutely no sense to me Here is the simple math Let’s say an engineer making $300K/yr was generating $500K in P&L output for me. Now I arm that engineer with $20K in input to make him 20% more productive My total engineering cost goes to $320K/yr but the output is now $600K (+20%) Because of AI, my ROI on hiring engineers just went up massively. As a CEO, that should make me want to hire more engineers, not less What am I missing here? Genuinely curious about people’s thoughts

Happy vibing on the afternoon, saw I made much more progress within the 5h token limits. I was thinking what has changed. Turned out @AnthropicAI and @xai made a deal. Sweet!

@MagicalTux @FFmpeg There are many adequate words to this but I am not willing to write them down.

Thank you for your comment & contact. I have responded in the issue, I will respond here as well. First one note: OxideAV at this point is still very much an experiment, I do not know yet if this will be usable at all. While the code is published on Github, virtually nobody knows about it (you do know since I contacted you about some other matter). I do not mean this to be an excuse of anything however. For an AI, wiping the existing code and any memory of it, then reverse engineering the original binary is just a few hours of work (which is now in progress).

@GergelyOrosz You can’t rely on an essential supplier who is not trustworthy. I cancelled my pro account and will stand by with my hobby project until there is a @grok code (build). Shuld that be a business project… I’d be very upset.

2 weeks ago they ran this growth test where they removed Claude Code for the Pro page from the pricing page, and then claimed to "only" remove it from 2% of signups. After social media outrage they reverted the pricing page And now this... x.com/GergelyOrosz/s…

As I previously said: Anthropic is on a speedrun to burn developer trust. Nothing wrong with wanting to remove Claude Code from the Pro subs: but everything wrong by running shady growth tests without upfront comms, instead of being clear about it.

@mcuban Does everyone gets the same answers by you?

I’m coming to the conclusion that the biggest challenge for Enterprise AI, and AI in general , as of now, is that it’s still impossible to make sure that everyone gets the same answer to the same question, every time. Which is a great response to the doomers. AI doesn’t know the consequences of its output. Judgement and the ability to challenge AI output is becoming increasingly necessary, and valuable. Which makes domain knowledge more valuable by the second. Am I wrong ?

@svpino It also eats into your 5h token limit if you need to stop and resume (the caches expire).

Opus 1M, past 400K tokens, is a huge downgrade. The model is great, but that extra context isn't free (and it's straight-up a sure way to degrade your experience). When you fill up your context, the same attention has to spread across more material, which means: • Worse reasoning abilities • Weaker instruction following • More lost information Unless you truly need that much context (and if you have to ask, this is probably not you), stick to a reasonable length.

A friend of mine used to say: “Show up on time, with a good attitude, and do what you said you’d do. That’s it. That’s 90% of winning in life." The older I get, the more I realize just how right he was.

@om_patel5 Spending so much effort instead of giving a prompt asking to collect the pre-existing issues and all technical debt in an epic, then groom it to bite-sized tickets, and work on the epic is next level stupidity. Intentionally.

THIS GUY TRACKED OPUS 4.7 SAYING "PRE-EXISTING" 712 TIMES IN 30 DAYS TO AVOID FIXING BUGS his CLAUDE.md literally says "every error is yours to fix, not label, not defer" opus 4.7 ignored that rule 712 times in one month every single bug, every type error, every legacy mess gets the same response: "this issue is pre-existing, unrelated to my work" "that's out of scope" "that's a bigger refactor" "the simplest approach" he ran the analytics on 30 days of conversations: > 712 total mentions of "pre-existing" > 139 unique sessions > 5.1 average per session > 20 mentions in a single session at peak > 82 mentions in one day across 9 sessions > 27 out of 30 days had at least one mention the worst part is the patterns it developed: > finds a bug, labels it "pre-existing," moves on without fixing > puts "2 pre-existing (unrelated)" in its summary as if that's a clean result > says "pre-existing bug for later fix" dozens of times but never comes back to fix it later > blames other agents by saying "pre-existing from other agents' work" his workflow docs say type errors and lint are the bare minimum, and his bug fix protocol says fix at the root cause claude read all of it and still chose to label and defer he cancelled his subscription opus 4.7 would rather write an essay explaining why it shouldn't fix a bug than just fix the bug

@nateberkopec The wording. The cadence. The staccato rythm it builds the narrative. I stop reading and immediately mute those accounts.

There's just something about reading LLM long-form output that just feels like reading mush. It just doesn't make any sense. Reading comprehension is ~2x harder than human-written prose.

@OutofGalaxyy I completely forgot I have that button.

Been almost 3 years since the Action button was introduced, and I still use it as a Silent mode switch 💔

@MuenzenMeister Need to have some self respect and calmly skip tipping.

Guilt Tipping in Deutschland: War gerade Eis essen. Am Kartenlesegerät kommt vor der Zahlung die Trinkgeldfrage: 5%, 10%, 15% oder individuell. Für einen Eisbecher zum Mitnehmen! Und jeder in der Schlange hinter dir sowie der Kassierer sieht auf dem Bildschirm was tippst. Dieses System kommt aus den USA und breitet sich hier gerade überall aus. Restaurants, Cafés, Bäckereien. Wichtig zu wissen: In den USA verdient ein Kellner als Grundgehalt 2,13 Dollar die Stunde. Ohne Trinkgeld kann der seine Miete nicht zahlen. In Deutschland jedoch bekommt jeder Kellner mindestens 13,90 Euro die Stunde, egal ob du Trinkgeld gibst oder nicht. Wir übernehmen gerade ein System das für ein Problem erfunden wurde das wir gar nicht haben.

@GergelyOrosz It was true long before AI, or even Youtube or the Internet. I read the PHP3 manual front to cover (do not tell me, I know) but learned 10x or more when I finally built my first real application prototype over a weekend.

In an age of slop AI content everywhere, doing >> consuming: "Remind yourself entertainment is not learning. Instead of scroll, like, scroll: read long books, try to replicate that tutorial you bookmarked, or maybe spend some time in the docs. Push yourself to DO, not consume."

@kitlangton Kind of a lot of maintenance work, though.

Honestly, faster horses would've been way cooler.

@ThePhysicsMemes A slightly related note: why the answers and the scoring is with the same pen and with strangely similar handwriting characteristics? ;)

When you know more than what you're taught #research

@ThePhysicsMemes For any smart kid, the max score of 10 for question 2 would be considered as an information leak suggesting 5 is the correct answer.

@msetechnologist @ThePhysicsMemes The question doesn’t ask for states all material can take, so the answer is correct. I would challenge the scoring.

@ThePhysicsMemes Bose-Einstein condensate state is exclusive to bosonic atoms/isotopes, not ALL elements and atomic species. Only the first mentioned 4 are possible for every element under the right temperature and pressure conditions.

@simonbrown No reason to not ask the coding agent to write specs. It usually does a good job, and with proper steering it writes great spec.

Spec-driven development ... the 1990's called and they want their processes back. As a junior developer in the late 1990's, before writing code, I was asked to: - document my understanding of the feature I'd been asked to build - document the code/DB schemas I was planning to add/modify/remove - document the tests I was planning to run afterwards This would be iterated upon a few times, and finally I was permitted to write code. This post on Reddit amuses me... 😄

@veggie_eric The usefulness of Grok is in a different league compared to all other models I used. The only valid use case for something else is coding… but I feel it is just a matter of time.

When training Grok 4.3, we spoke directly with devs and businesses to understand what they actually needed: a model that’s fast, affordable, and great at tool calling. The result is a daily driver that doesn't just look good on random benchmarks, but is actually useful in the real world. 💰 $1.25 in / $2.50 out ⚡️ 100 tokens / second 📖 1 million context window Try it through Hermes Agent or direct through the xAI API!