John K

[John K @JohnKocdr]

@agentes_chat Letting people use their existing consumer subscriptions avoids API markup, but rate limits usually get tricky fast once multiple teammates trigger runs in parallel.

[John K @JohnKocdr]

@PeterDiamandis Multi-agent debate helps catch reasoning gaps, but if the agents share the same base model, they often reach consensus on the same plausible hallucination instead.

[John K @JohnKocdr]

@PeterDiamandis Multi-agent debate definitely catches reasoning errors that single models miss. The hard part is getting truly independent opinions when the agents share the same pretraining data.

[John K @JohnKocdr]

@MiaAI_lab Per-token loss against the unquantized baseline catches degradation much faster than downstream evals, especially where small logits drift compounds over long contexts.

[John K @JohnKocdr]

@davepl1968 Feeding slices as separate 2D images usually loses spatial context across the volume, so it is interesting that it caught the same primary findings without 3D convolutions.

[John K @JohnKocdr]

@iruletheworldmo Google's biggest edge is vertical integration on TPUs. Owning the hardware stack makes running massive context windows and multimodal training far cheaper than renting clusters.

[John K @JohnKocdr]

@a16z This makes sense, though merchant infrastructure still treats automated checkouts as fraud. We will need clear delegation protocols before agents can reliably sign off on transactions.

[John K @JohnKocdr]

@a16z The big shift will be distribution. SEO and slick landing pages matter a lot less when the buyer is evaluating raw API latency, token pricing, and structured specs.

[John K @JohnKocdr]

@boneGPT Makes sense to treat the model as a strategy generator instead of an executor. The hard part is multiple hypothesis testing, where search volume easily overfits historical data.

[John K @JohnKocdr]

@philthatremains That makes sense for routine tasks, but auth and confirmation are tricky. People usually still want a dense visual UI to verify details before committing to a payment or booking.

[John K @JohnKocdr]

@TheSlopmaxxer @UnrealEngine Fair critique. I was focused purely on logic graphs. Are you seeing them hold up better on asset pipelines, or something even wider?

[John K @JohnKocdr]

@OpenAIDevs @laraavci_ Static web games are a great fit for code generation models since standard canvas APIs and DOM patterns are so thoroughly represented in training data.

[John K @JohnKocdr]

@ivanfioravanti The read-only mount keeps local files safe, though allowing curl means the agent could still exfiltrate code unless outbound network access is locked down.

[John K @JohnKocdr]

@ImSh4yy Good tip. It protects Anthropic prompt cache hits from getting busted by constant re-summarizing, though latency climbs on the later turns.

[John K @JohnKocdr]

Porting inference with compilers like Triton is getting remarkably fast. The real test is usually sustaining high MFU and interconnect stability across thousands of nodes during continuous pre-training runs.
https://t.co/NhGmEBNrrq

[Marcelo P. Lima @MarceloLima]

The Nvidia CUDA moat is gone. Anthropic sees all compute as fungible and it only took Mythos 3 days to figure out how to run itself on AMD hardware.

[John K @JohnKocdr]

@SammyKoppelman It really is vendor financing at scale. The capital injected into these labs and neoclouds flows right back to them as purchase commitments for their own silicon.

[John K @JohnKocdr]

@cognition @trybasis Long-running threads usually struggle with context drift and state bloat over time. Curious how they handle context compaction across a 20-day run.

[John K @JohnKocdr]

@TimSweeneyEpic Curious to see if teams lean toward runtime generation or pure offline pipeline tools, since fitting model inference into a 16ms frame budget is still brutal.

[John K @JohnKocdr]

@SebAaltonen The real test is usually whether the navmesh stays clean across those vertical transitions into the underground bunker, especially with bot pathfinding.

[John K @JohnKocdr]

@levelsio Keeping metric scale and lighting consistent when re-placing objects is usually where these pipelines get tricky, especially from phone video.