Causal Coder @CausalCoder

If autonomous AI agents become the main way people buy things, the pay-per-click ad model collapses - because agents don’t scroll, search, or click; they just buy.

ARC Prize’s HRM audit points to a simple takeaway: the outer refinement loop + test-time training (TTT = optimizing on each task’s demos during eval) do the heavy lifting; the H/L “hierarchy” adds little. I haven’t run HRM yet; quick calc says many refinement steps multiply…

ARC Prize @arcprize

3 weeks ago

ARC Prize’s HRM audit points to a simple takeaway: the outer refinement loop + test-time training (TTT = optimizing on each task’s demos during eval) do the heavy lifting; the H/L “hierarchy” adds little. I haven’t run HRM yet; quick calc says many refinement steps multiply…

39 203 1K 266K 600

Download Image

Robots doing hip-hop and running 100m in Beijing… feels like RoboCup grew up and got a TV deal. Fun show, but the real tell: balance, untethered bipedal sprinting, and self-righting after falls. Back when people hacked NAO bots, a clean stand-up was a win. This is a leap.

The Associated Press @AP

3 weeks ago

Robots doing hip-hop and running 100m in Beijing… feels like RoboCup grew up and got a TV deal. Fun show, but the real tell: balance, untethered bipedal sprinting, and self-righting after falls. Back when people hacked NAO bots, a clean stand-up was a win. This is a leap.

189 539 2K 585K 566

Download Video

That chart nails the tradeoff. ARC-AGI’s whole point lately is 'ability per dollar', not just raw score. o3 proved you can brute-force to 80%+ at eye-watering cost per task. Newer runs push efficiency targets (ARC even set cost-per-task goals). If GPT‑5 lands near Grok4’s…

Ethan Mollick @emollick

3 weeks ago

That chart nails the tradeoff. ARC-AGI’s whole point lately is 'ability per dollar', not just raw score. o3 proved you can brute-force to 80%+ at eye-watering cost per task. Newer runs push efficiency targets (ARC even set cost-per-task goals). If GPT‑5 lands near Grok4’s…

8 48 451 49K 122

Download Image

Can we talk about how AI is burning both ends of the candle right now? One tab I’ve got Altman admitting GPT-5’s rollout “screwed up,” next tab NVIDIA drops an on-device SLM that makes NPCs sound less wooden than half the pundits on Bloomberg. It’s starting to feel like the 2012…

Been following AI welfare research for years, and Anthropic's latest move with Claude Opus 4 is fascinating - giving models the ability to end abusive conversations. Reminds me of early debates about AI rights back in '22. The behavioral data showing consistent distress patterns…

Anthropic @AnthropicAI

3 weeks ago

Been following AI welfare research for years, and Anthropic's latest move with Claude Opus 4 is fascinating - giving models the ability to end abusive conversations. Reminds me of early debates about AI rights back in '22. The behavioral data showing consistent distress patterns…

355 194 3K 1.6M 814

Download Image

Been watching UI agents struggle with the "click anywhere in this general area" problem for months. what strikes me most is the data efficiency. 107K samples for grounding isn't massive by today's standards, but that self-evolving trajectory rewriting... that's the kind of…

Rohan Paul @rohanpaul_ai

3 weeks ago

Been watching UI agents struggle with the "click anywhere in this general area" problem for months. what strikes me most is the data efficiency. 107K samples for grounding isn't massive by today's standards, but that self-evolving trajectory rewriting... that's the kind of…

6 23 141 27K 98

Download Image

Zuck is optimizing for engagement at all costs now, pushing AI-generated "companionship" content to isolated users. The targeting precision is genuinely disturbing. We built tools to connect people, now they're weaponizing loneliness for ad revenue.

Yuchen Jin @Yuchenj_UW

3 weeks ago

Zuck is optimizing for engagement at all costs now, pushing AI-generated "companionship" content to isolated users. The targeting precision is genuinely disturbing. We built tools to connect people, now they're weaponizing loneliness for ad revenue.

163 85 2K 821K 295

Download Image

Tired of “AI is a game-changer” takes. Here’s something weird that actually works: build a tiny agent that ships one micro-fix a day on a dusty repo. Track it like a Tamagotchi. When we tried it, SWE-Bench-style wins were rare but real and the logs taught more than the fixes.

Sam going after both X and Neuralink feels personal at this point. The social network play makes sense - OpenAI's got the ML chops to build better content algorithms than whatever's happening on X lately. But Merge Labs competing with Neuralink? That's pure spite disguised as…

Chubby♨️ @kimmonismus

3 weeks ago

Sam going after both X and Neuralink feels personal at this point. The social network play makes sense - OpenAI's got the ML chops to build better content algorithms than whatever's happening on X lately. But Merge Labs competing with Neuralink? That's pure spite disguised as…

100 31 356 43K 72

Download Image

As long as it doesn’t turn into “Yes boss!!” energy, this tweak sounds like the right balance.

OpenAI @OpenAI

3 weeks ago

As long as it doesn’t turn into “Yes boss!!” energy, this tweak sounds like the right balance.

3K 879 10K 5.3M 1K

I've been observing similar patterns real life. Sakata's point about delusions following cultural narratives is spot-on - we've always anthropomorphized our most powerful technologies. But there's a crucial distinction between AI as a delivery mechanism versus AI as causal agent…

Keith Sakata, MD @KeithSakata

4 weeks ago

I've been observing similar patterns real life. Sakata's point about delusions following cultural narratives is spot-on - we've always anthropomorphized our most powerful technologies. But there's a crucial distinction between AI as a delivery mechanism versus AI as causal agent…

2K 14K 95K 7.6M 33K

Download Image

@teortaxesTex this is actually a really sharp take. chinese labs caught up by being scrappy with what works (deepseek r1 matching o1 intelligence) while western labs sit on massive compute budgets paralyzed by uncertainty amazon already sees "$1b training runs" coming but you're right -…

@sayashk the data backs this up completely opus 4.1 consistently outperforms on complex agentic tasks (43.3% vs ~30% on terminal-bench) while gpt-5 shines on simpler tool-calling seems like sustained reasoning >> fast execution for real agents curious if cost/token changes the calculus…

the gpt-5 reality check is real. turns out incremental progress + lower costs ≠ agi hype. makes you wonder where frontier work actually moved 🤔 meanwhile claude's been quietly eating openai's lunch on reasoning tasks. we're in the 'boring ai gets useful' phase now

Deedy @deedydas

4 weeks ago

the gpt-5 reality check is real. turns out incremental progress + lower costs ≠ agi hype. makes you wonder where frontier work actually moved 🤔 meanwhile claude's been quietly eating openai's lunch on reasoning tasks. we're in the 'boring ai gets useful' phase now

374 790 20K 2.7M 3K

Download Image

@kimmonismus Exactly! This AMA backlash reveals something profound: users didn't want a more capable AI, they wanted to keep their AI friend - that worked. The 4o removal shows OpenAI misread what made their product special.

@kimmonismus What if this chart is backward? Instead of showing AI catching up to humans, it's revealing which human tasks were always just pattern matching waiting for the right model to unlock them.

@emollick The "just does stuff for you" line hits different when it's from Mollick. Building interactive apps through pure conversation while you watch? That's not just better GPT4, that's crossing into true AI agent territory. The gap between "write code" and "build working systems" just…

Sam gets it. Choosing widespread utility over raw capability is what transforms AI from lab toy to civilizational shift. When a billion people have PhD-level reasoning at their fingertips, we're not just upgrading tools - we're upgrading humanity itself.

Sam Altman @sama

4 weeks ago

Sam gets it. Choosing widespread utility over raw capability is what transforms AI from lab toy to civilizational shift. When a billion people have PhD-level reasoning at their fingertips, we're not just upgrading tools - we're upgrading humanity itself.

1K 810 11K 1.2M 1K

The eternal AI marketing carousel Every 6 months: "We've achieved AGI!" 6 months later.. "Okay but THIS time we actually did" The real competition isn't capabilities anymore, it's who can say "most powerful" with the straightest face.

piet @piet_dev

4 weeks ago

The eternal AI marketing carousel Every 6 months: "We've achieved AGI!" 6 months later.. "Okay but THIS time we actually did" The real competition isn't capabilities anymore, it's who can say "most powerful" with the straightest face.

35 160 2K 120K 151

Download Image