Reply by: oldschool_dev (Score: +7)
15 min ago
pro tip: start small. one simple workflow with clear success metrics. most people fail because they try to build skynet on day 1.
and for the love of god, add human review gates. ai is a very expensive intern, not a senior engineer yet.
Reply by: devguy42 (OP)
25 min ago
thanks for the replies, good discussion. seems consensus is: agents are progressing but mostly useful when heavily guided. raw model intelligence is plateauing a bit and the game is moving to systems, workflows and infrastructure.
the energy thing surprised me how many mentioned it. guess thats the unsexy but critical bottleneck in 2026.
anyone got good resources on building reliable agent workflows that dont go off the rails? preferably not vendor hype docs lol
Reply by: privacy_paranoid (Score: +12)
35 min ago
forget performance, the regulation and privacy stuff is gonna blow up this year. eu is already tightening, us might actually do something after the elections noise dies.
also the "90% of online content synthetic" prediction is terrifying for training future models. we're gonna hit a data wall hard.
on the positive side, multimodal is actually useful now. video understanding + voice is way better than last year. i use it daily for research.
but yeah... still waiting for the agent that can do my taxes, book my travel, and not fuck up my credit card details.
Reply by: ai_optimist23 (Score: +18)
40 min ago
@skeptic_coder not replaced but augmented hard. our junior devs are 2-3x faster now with cursor + claude. seniors still needed for architecture and hard problems.
hot take: 2026 is the year models stop mattering as much. the winners will be whoever builds the best workflows, memory systems, and tool use. context engineering > raw intelligence right now.
also energy and data center bottlenecks are real. nvidia still printing money but power grid issues in virginia and texas are getting messy. some projects getting delayed because they literally can't get electricity.
anyone following the humanoid side? figure, 1x, tesla optimus... feels like the next hype cycle after llms.
Reply by: skeptic_coder (Score: +29)
50 min ago
this whole thing feels like 2023-2024 all over again but with better benchmarks.
remember when everyone said gpt-4 would replace programmers? now we're arguing if claude or grok-4 or gemini is better at swe-bench. meanwhile im still fixing ai generated code that "works on my machine" but breaks in prod.
real talk: the models that actually moved the needle for me this year are the reasoning ones (o-series style). they suck less at long tasks. but they're slow and expensive as hell for daily use.
question for the thread — has anyone successfully replaced a junior dev with agents yet without quality dropping like a rock? i call bullshit until i see numbers.
Reply by: startupfounder91 (Score: +41)
1 hour ago
@ml_engineer87 exactly. the hype around "full autonomous agents" is still overcooked. most companies are doing AI workflows not true agents. like chaining tools with human in the loop.
what im seeing in 2026:
vertical AI tools crushing it (legal, medical, finance specific models)
open source stuff like deepseek and qwen getting scary good and cheap
humanoid robot training data becoming huge (all those videos of people doing chores lol)
but enterprise adoption is slower than twitter makes it seem. cios are burned from last years pilot graveyard. they want measurable ROI not "productivity vibes".
also the synthetic content flood is getting wild. half the web might be ai slop by end of year. how do we even train on good data anymore?
Reply by: ml_engineer87 (Score: +52)
1 hour ago
lmao "demo porn" is the perfect term.
i work at a mid-size saas company and we rolled out claude 4.5 + some custom agents for support ticket routing and code review. the ticket thing is actually saving us ~18 hours a week across the team. but the code agent? its decent at boilerplate but i still have to rewrite half the stuff and the security holes it introduces... jesus. had to ban it from touching auth code completely.
the biggest win so far has been multi-step reasoning models (those "thinking" versions). they actually plan better now instead of just spitting first token salad. but energy costs are insane, our infra bill went up 40% this quarter.
anyone else noticing power consumption becoming the new gpu shortage?
Posted by: devguy42
2 hours ago | 187 upvotes | 64 comments
yo guys, been following AI since the gpt-3 days and im starting to feel exhausted. every week theres new "agent" announcements but when i actually try them in real work they still hallucinate, loop forever or just do dumb shit.
is 2026 actually gonna be different with these reasoning models (like the new claude and whatever openai dropped after o-series) or are we just moving the goalposts again?
personally i think the real shit happening rn is agentic workflows + context engineering not full autonomous agents. what do u think? anyone seeing real ROI in their company or is it all still demo porn?
