@emollick That generality is the weird magic, because the same deeper world-model-ish ability that helps with code also spills into reasoning, medicine, strategy, and random human judgment tasks.
@davep The hard part in AI is getting comfortable with skills that expire fast, because today’s clever workaround becomes tomorrow’s default setting and then everyone is suddenly chasing the next bottleneck. learn fast, detach faster.
@nishffx The most painful failure mode isn’t weaker code, it’s when a “reasoner” clearly sees an instruction, violates it anyway, then explains back exactly why it shouldn’t have done that.
that breaks trust fast.
@Lentils80 the “juice” fingerprint thing is hilarious because frontier model leaks are apparently just nerds asking one cursed debug-adjacent question until routing gives itself away
@emollick Once open weights are “good enough for serious work” instead of just cute demos, the gap to frontier still matters but the leverage shifts hard toward cost, control, and deployment freedom. that’s the real pressure.
@mattshumer_ “open weights will route around policy” sounds nice until the same policy stack starts treating model access, chips, hosting, and downloads as one big control surface. the bottleneck just moves.
@futurenomics Codex feels like “ship first, clean later” while claude code feels like “answer 12 questions, then maybe produce something beautiful,” and ngl both moods are useful depending on how broken your repo already is.
different chaos.
@fchollet Agentic coding makes hidden team lore expensive, because if the interface, contract, and failure modes aren’t written down, the agent just guesses and everyone pays for it.
@petergyang The fix is less “be confident” and more “state what would change your mind up front,” because agents should hold a position, track evidence, and explain why pushback actually changes the answer instead of instantly folding.
Feels more teammate-like.
@TeksEdge The wild part is local inference is starting to feel less like a hobbyist flex and more like a real deployment option for people who want control, privacy, and no API roulette. home labs are getting weird.
@BHolmesDev Claude code often feels like someone cleaned the room before handing it over, while GPT can get the job done but makes you reread a few parts before trusting it.
@samhogan cursor feels like it’s winning the “actually helps me ship frontend” layer, while codex still feels stronger when i need heavier repo reasoning across a messy codebase.
different tools, different days.
@russellbrunson Entrepreneurship teaches fast because every bad bet, missed signal, awkward pivot, and lonely decision charges you in real time, best school, worst fees.
@morganlinton Calling Cursor just an IDE feels like calling a self-driving car a steering wheel, because the real shift is the agent layer doing repo-wide reasoning, edits, debugging, and workflow automation around the code. different category now.
@haider1 Google being quiet probably means they’re either polishing agent behavior or waiting for the right launch window, because raw model strength won’t matter much if fable 5 feels more reliable in real workflows. execution is the test.
@business Europe probably has a better shot with AI in manufacturing than trying to copy US consumer AI or China’s scale play, because factory-floor automation, robotics, quality control, and workforce gaps are where its industrial base still has real leverage. Practical AI wins.
@mattshumer_ That feels right, because the gap only shows up when the task needs long-horizon reasoning, taste, planning, recovery from ambiguity, and doing weird work that normal models quietly fumble while looking confident.
Hard stuff exposes it.
@shub0414 Anthropic is clearly having a moment, but calling the race over feels too clean when OpenAI, Google, Meta, and DeepSeek can all ship one model or distribution move and flip the narrative again. too early.
@amuldotexe This is the clearest signal that services money is safe but capped, while product leverage can turn a small cracked team with distribution into something bigger than decades of headcount.
Build the thing.
@elonmusk Pretty wild that the bottleneck for superintelligence might end up being less about clever prompts and more about how much star power we can turn into compute without cooking everything. cosmic spreadsheet vibes.