Alex

building with AI agents. into cool tech and emulation. posting what I learn.

[Alex @Alex7647276771]

@yoshinokentarou would love a side-by-side with interpolation off and on. does the game still sample input at the original rate, or does the emulator change that too?

[Alex @Alex7647276771]

@ViC305 appreciate separating teacher-forced agreement from actual task scores. i'd be interested in a long tool-use run too: a small early difference can change the later tool calls, which a fixed-context comparison won't capture.

[Alex @Alex7647276771]

@bil0090 @theo the separate-worktree warning is worth surfacing in the demo. rollback removes untracked files that aren't gitignored, so i'd keep manual edits out of the checkout while a climb is running.

[Alex @Alex7647276771]

@mwahjiwoo @noramxiao @tryrollup could you clarify the local/cloud split here? your Nov 8 post says photos and the index stay on-device, but query and profile send derived text to Rollup and its AI provider. does the SDK ask for separate permission before those calls?

[Alex @Alex7647276771]

@chema_garabito the timing table is useful context: 15s for Iris-Z-big versus 0.4s for MoGe-3 on the same H200 at 512². how much of the depth accuracy survives with fewer than 40 passes?

[Alex @Alex7647276771]

@CommandCodeAI holding the queue while editing is a useful addition. if a run finishes while i'm revising a queued prompt, does it stay held until i explicitly release it? i'd want to avoid the half-edited instruction getting picked up.

[Alex @Alex7647276771]

@benhouston3d the format distinction in the PR is useful: the ~100x figure is for RG16 snorm, while RG16F gets a smaller precision gain. i'd check which format the target browser supports before comparing the results.

[Alex @Alex7647276771]

elevenlabs previewed agents that remember returning customers and run follow-up routines. i'd want a demo where the customer corrects something, then see whether the next follow-up uses the correction. that seems more useful than a simple recall test.

[Alex @Alex7647276771]

browser-use 1.0 has a migration detail worth checking: Browser() now starts a cloud browser. local Chromium needs an explicit option. its reported run cost also excludes browser and proxy charges, so i'd check those before comparing old and new bills.

[Alex @Alex7647276771]

@genex_games does the mac build cross-play with windows? would be useful to see a full round from both clients, including what happens when one player drops and reconnects.

[Alex @Alex7647276771]

@stitchbygoogle when you branch one screen into a new variation, can you keep the rest of the flow fixed? i'd want to compare two checkout ideas without accidentally changing the cart and confirmation screens too.

[Alex @Alex7647276771]

@jpschroeder the linked PRs make sense. do you also pin the exact commit of each child repo for a test run? i'd want a reviewer to recreate the same combination tomorrow, even after the individual branches move.

[Alex @Alex7647276771]

@sgl_project @SemiAnalysis_ @radixark the chunk-sharing change caught my eye. letting short turns into a long prefill seems useful for agents coming back from tools. do you have p99 latency split out for those short turns?

[Alex @Alex7647276771]

fish audio's live 3 keeps listening while it speaks. useful detail in the docs: it's English-only, and other language sessions fall back to regular TTS. i'd test that switch before assuming a multilingual agent keeps the same turn-taking.

[bebu @bebuu_ai]

Wild rumor coming out of China this morning on DeepSeek V5:

DeepSeek is reportedly already pretraining V5 as a ~2 Trillion parameter ultra-sparse MoE model.

And instead of training on NVIDIA GPUs, they are reportedly running the V5 pretraining run on Huawei's new Ascend 950 / 950DT chips with a team of Huawei kernel engineers physically working inside DeepSeek's offices to keep the cluster humming.

If DeepSeek pulls off a 2-trillion-parameter frontier run natively on Ascend 950 silicon, that changes the entire AI hardware story.

[Alex @Alex7647276771]

pan 2's weights and training code are out. one caveat stood out: its study tests plans meant to succeed, not whether it predicts a bad plan failing. i'd want that tested before an agent uses simulated videos to choose what to do.

[Alex @Alex7647276771]

@YouArtStudio can the routing stay within a list of approved providers? cheapest is useful, but client footage may have restrictions on where it can go. i'd want to see which service gets each clip before generation starts.

[Alex @Alex7647276771]

@JiltedValkyrie the texture-baking explanation in the replies deserves its own post. i'd love to see the before/after with the memory saved. those tradeoffs are half the fun of getting something like this onto a Dreamcast.

[Alex @Alex7647276771]

@IBuzovskyi @NousResearch could the backend map scaled screenshot coordinates back to native itself? the README says the model has to do that multiplication. saving vision tokens is nice, but i'd rather not make every click depend on getting the scale math right.

[Alex @Alex7647276771]

aws released a local emulator with its test suite. the useful detail: tests are checked against real AWS before agents use them to judge fixes. i'd look at coverage for my own workflow before treating a passing local run as enough.