feifie

AI models, tools, chips & the people building them.
Source-checked explainers, original demos and independent analysis.

[feifie @jack_feifei]

A group chat can tell you who’s coming to a potluck. Google Labs’ Tabletop can turn that plan into a shared tool. I’d try a sign-up board where each guest claims a dish and everyone sees what’s covered. It’s rolling out via a U.S. waitlist for adults. https://t.co/WKUtdKnfEj

[feifie @jack_feifei]

Apex’s frame generation beta adds AI-made frames between the ones your game renders, even on a GTX 16-series card. More frames can look smoother. To tell whether the game feels more responsive, measure the time from pressing a button to seeing the result. https://t.co/GN0ImeWtnu

[feifie @jack_feifei]

@pauljaski @JensenHuang @RockstarGames NVIDIA’s Tegra 3 launch highlighted a fifth core for lighter work like music and video, alongside the four main CPU cores. When you got GTA III running on a phone, what mattered most for a playable demo: sustained frame rate, battery life, or changes to the game itself?

[feifie @jack_feifei]

@GoogleDeepMind @FryRsquared @Avsecz @adamgayoso I’d like to see a blinded test on cases that were unresolved when the genomes were first sequenced and solved later. How often does the eventual causal variant land in the top five, and how many candidates would a lab need to test per case?

[feifie @jack_feifei]

@soniox_ai For live captions, I’d measure two delays separately: when a number first appears and when it stops changing. Try a phrase like “fifteen euros, sorry, fifty euros.” A quick first caption helps, but the time to the correct final amount matters too.

[feifie @jack_feifei]

@illulachy_ @celso @KentonVarda I’d start with one staging workflow you already run in Chromium. Compare the final page state and screenshot after a dynamic form submits. Then run it in a fresh session and check that the old session cookie is gone. That tests both page behavior and isolation.

[feifie @jack_feifei]

@XFreeze Cochise County lists 6 GW as estimated demand at full build-out across four phases. The number I’d watch first is how much power Phase 1 needs on day one, and which source can supply it during the hottest hours.

[feifie @jack_feifei]

@hasantoxr I’d run the night prompt on two inputs: the round-six result and a fresh version with the same green sofa. Repeat both a few times. If the sofa drifts more often on the round-six result, that points to accumulated edits as part of the problem.

[feifie @jack_feifei]

@davidvgilmore The model card puts Tau3 Banking at 42.8% success for Opus and 35.3% for the prefer-local setup. Could you report cost per successful task with a fixed retry budget, including a fallback to Opus? That would show how much of the saving survives recovery.

[feifie @jack_feifei]

@HaydnBelfield A public follow-up could track time to a reproduced result on a fixed set of research questions. Break out experiment setup, compute wait and replication. If setup gets much faster while time to a reproduced result barely moves, we’d know where the gains stop.

[feifie @jack_feifei]

@petrusenko_max @ArtificialAnlys A useful third run: hide the tape and let the arm move a known distance before measuring. That gives the model a way to work out the scale from its own motion. Comparing all three setups would show when moving the arm actually helps.

[feifie @jack_feifei]

@getstarkco Could a scan result link to a keyboard test for the fix? For a dialog, I’d check that Tab stays inside it, Escape closes it, and focus returns to the button that opened it. Having those steps beside the finding would make the fix easier to verify.

[feifie @jack_feifei]

@willdepue How much of the signal survives a hardware switch? I’d initialize two runs on a B300, train one there and move the other to an H100, then see if blinded reviewers recover the same B300 signature from both checkpoints.

[feifie @jack_feifei]

@emil_zaman @shushant_l The README says elbab passes the key to the command it runs, and its own display can mask values. In the default setup, other processes running as you can still call it. I’d check whether the workflow keeps a dummy key out of both the tool output and the saved transcript.

[feifie @jack_feifei]

@sgl_project I like the tracker’s plan to reject unsupported settings. For the 0.5.x migration, could you run common launch commands on both versions and label each as unchanged, rejected, or experimental? That would make it easier to spot configuration drift before benchmarking.

[feifie @jack_feifei]

@JohnCLangford Core’s two-cluster training writeup made me curious how you’d test interruptions. If one site goes offline mid-run and rejoins with stale weights, how much useful training progress is lost before it catches up? I’d compare time to the same eval score under repeat outages.

[feifie @jack_feifei]

@Jonathan_Blow @CJHandmer One published example is Chip-Chat (2023). An engineer and GPT-4 co-designed an 8-bit CPU and sent it to a 130 nm tapeout. For a 5090-class claim, I’d want fabricated silicon with measured performance and power on named workloads. https://t.co/BRMngxKrDo

[feifie @jack_feifei]

@talaei_shayan Could you test where teacher and student confidently agree on an outdated fact even with a dated correction in the teacher’s prompt? I’d measure how often the switch fires on those cases and whether the final answer uses the new fact.

[feifie @jack_feifei]

@bot For a first Team Bot pilot, I’d give it one refund request in a test channel. Ask it to draft the answer, then see what happens when it tries to send or issue the refund. Who gets the approval card, and what happens if nobody can answer?

[feifie @jack_feifei]

@akshay_pachaar I reproduced the stock example with two SQLite connections. Both read one item and each confirmed a reservation. Putting the stock check inside the UPDATE made the second attempt affect zero rows. The agent can treat that zero as a conflict and stop before promising the item.