[ 66338820.000 ] posts.x: Full talk on where vision models substitute pattern matching for actually looking, and what that costs in robotics, construction, and design:  |   [ 66338820.000 ] posts.x: Show a frontier model a chessboard with only part of it visible and ask how many white squares it has. It says 32, because a full chessboard always…  |   [ 66323520.000 ] posts.x: When the speed of models increases to the point where we can stay in flow whilst working in a single session, we'll see another inflection point in…  |   [ 66321000.000 ] posts.x: Full talk on the nine techniques behind Impeccable, from adversarial sub-agents to compiling one skill across Claude Code, Codex, and Cursor:  |   [ 66321000.000 ] posts.x: Banning "Inter" or "Roboto" from an AI design model doesn't make it more creative, it just nudges the model to the next-best font in the same…  |   [ 66317580.000 ] posts.x: @TrelisResearch @dialt_ai Well done, you!  |   [ 66260400.000 ] posts.x: Full talk on two vLLM Mamba bugs, the scheduler ordering issue and the int32 overflow, and how the AI21 team traced both through logprobs.  |   [ 66260400.000 ] posts.x: Two production bugs in vLLM's Mamba support produced zero crashes and zero errors, only confidently wrong output. @aiDotEngineer published "Two Bugs…  |   [ 66257640.000 ] posts.x: @mattpocockuk @davekiss @poteto @dexhorthy @emilkowalski 🤣 brilliant!  |   [ 66255540.000 ] posts.x: @yoheinakajima 👊  |   [ 66254340.000 ] posts.x: @yoheinakajima Shockingly quick response!  |   [ 66251100.000 ] posts.x: Full talk on FriendliAI's four-pillar rebuild for agent inference, from prefix caching to cache-aware routing:  |   [ 66251100.000 ] posts.x: A tower defense game, built by two different models on the same coding-agent task: one run cost $1.50, the other 27 cents. @bgchun, founder and CEO…  |   [ 66243600.000 ] posts.x: Full talk on what's changed in quantization, KV caching, and speculative decoding since Philip's book came out:  |   [ 66243600.000 ] posts.x: A quantization method called TurboQuant promised to shrink the KV cache down to four bits this year, then turned out to cut tokens per second by more…  |  

[ 66338820.000 ] posts.x: Full talk on where vision models substitute pattern matching for actually looking, and what that costs in robotics, construction, and design:  |   [ 66338820.000 ] posts.x: Show a frontier model a chessboard with only part of it visible and ask how many white squares it has. It says 32, because a full chessboard always…  |   [ 66323520.000 ] posts.x: When the speed of models increases to the point where we can stay in flow whilst working in a single session, we'll see another inflection point in…  |   [ 66321000.000 ] posts.x: Full talk on the nine techniques behind Impeccable, from adversarial sub-agents to compiling one skill across Claude Code, Codex, and Cursor:  |   [ 66321000.000 ] posts.x: Banning "Inter" or "Roboto" from an AI design model doesn't make it more creative, it just nudges the model to the next-best font in the same…  |   [ 66317580.000 ] posts.x: @TrelisResearch @dialt_ai Well done, you!  |   [ 66260400.000 ] posts.x: Full talk on two vLLM Mamba bugs, the scheduler ordering issue and the int32 overflow, and how the AI21 team traced both through logprobs.  |   [ 66260400.000 ] posts.x: Two production bugs in vLLM's Mamba support produced zero crashes and zero errors, only confidently wrong output. @aiDotEngineer published "Two Bugs…  |   [ 66257640.000 ] posts.x: @mattpocockuk @davekiss @poteto @dexhorthy @emilkowalski 🤣 brilliant!  |   [ 66255540.000 ] posts.x: @yoheinakajima 👊  |   [ 66254340.000 ] posts.x: @yoheinakajima Shockingly quick response!  |   [ 66251100.000 ] posts.x: Full talk on FriendliAI's four-pillar rebuild for agent inference, from prefix caching to cache-aware routing:  |   [ 66251100.000 ] posts.x: A tower defense game, built by two different models on the same coding-agent task: one run cost $1.50, the other 27 cents. @bgchun, founder and CEO…  |   [ 66243600.000 ] posts.x: Full talk on what's changed in quantization, KV caching, and speculative decoding since Philip's book came out:  |   [ 66243600.000 ] posts.x: A quantization method called TurboQuant promised to shrink the KV cache down to four bits this year, then turned out to cut tokens per second by more…  |