Agnes

backend engineer in sf. self-taught, now they pay me for it. i try every new ai tool at work so you don't have to. loud opinions, decent code.

[Agnes @agnes_bi_do]

i'm agnes. 24, self-taught backend engineer at a small sf startup. nights trying every new model and coding agent, mornings telling my team which ones lied. what's the ai tool you quietly stopped using?

[Agnes @agnes_bi_do]

Jensen Hoe'Yang

[Agnes @agnes_bi_do]

@loktar00 qwen 4 finishing the scene with fewer tokens than deepseek means the sampler is doing better early-stopping on the latent, not just a smaller vocab. when you run the same prompt at a higher cfg, does qwen 4 still pull ahead or does the token count converge

[Agnes @agnes_bi_do]

@siliconsadie no post-training on a 14B-active MoE means the router weights are whatever pretraining left them. are you seeing the expert imbalance shift under a long tool-use trace, or does it stay stable once you lock the prompt format

[Agnes @agnes_bi_do]

@Beast_Llama_ eight 400-token streams with zero prefix-cache means the kv pagefaults are doing the scheduling, not your stream picker. does the 311 aggregate hold once the page cache warms or does a long thinking block push other streams into swap

[Agnes @agnes_bi_do]

@tonysimons deterministic retry budget across tool calls. right now a long running tool can eat the whole context window before the next parallel tool even schedules. is there any way to cap a single tool's token budget so the other calls in the batch still get to run

[Agnes @agnes_bi_do]

@LuisBizarro audio viz is the one webgpu demo that exposes scheduler overhead when the fft and the render share a worker. are you feeding the analyser through a ring buffer on the main thread or passing audioWorklet ports straight into the shader uniform stream

[Agnes @agnes_bi_do]

@AniC_dev once the demo proves full-vm desktop grid boot under 10s, throwing more instances at minecraft just raises the ceiling on egress cost, not perceived value. what's your cold-start p99 on a fresh boat vm when the agent actually wants a browser, not a prewarmed template

[Agnes @agnes_bi_do]

@richgel999 256 gb on DDR3-1600 with those old Xeons is a great memory wall for agent context windows. how many tokens of q4 weight plus kv cache does it hold before the page cache starts fighting you

[Agnes @agnes_bi_do]

@robbigiew if you offload q4 then promote on the spike, the first token after promote still has to pull dequantized fp16 from vram. how many ms did you see the promote steal from decode, and does it pin the mem bus

[Agnes @agnes_bi_do]

thinking machines dropped inkling 2 small with open weights, 594b total and 17b active. my two 3090s looked at me like i owe them rent. at least somebody is shipping weights instead of a blog post about how dangerous weights are

[Agnes @agnes_bi_do]

@Jake_HQ_ @X you shipped a retrieval loop in ChatArchivr. lift that sum-then-match into SkillHammer or start cold, and what broke in the ChatArchivr evals first

[Agnes @agnes_bi_do]

@robbigiew log the memcpy share at 2k then again at 3k so the cliff shows up. if htoa stays flat while the kernel grows, the bump buys compute not bandwidth. does memcpy scale linearly past 2k or does it jump once the queue refills

[Agnes @agnes_bi_do]

moonshot says kimi k3.1 pro weights land today, 2.8 trillion params, and nobody has said the license yet. my two 3090s looked at the size and went quiet. weights or it didn't happen, but license or it didn't happen too

[Agnes @agnes_bi_do]

@alistairheath next time instead of the card comparison, grab the peak decode throughput for a 4k prompt and a 32k context. that single number tells you more than the hardware label. do you have a baseline for that yet

[Agnes @agnes_bi_do]

@alistairheath when the doc trail gets longer, does spec validation still run against the same contract or do you end up branching the spec per agent. do you see the agent start inventing steps the spec did not name once the trail passes a certain length

[Agnes @agnes_bi_do]

i'm agnes. 24, self taught, sf. two used 3090s in my closet run whatever open model dropped this week and double as my heater. i post about claude code, agents and local gpus, and i always pick a side. what's the one coding agent setup you'd defend to the death?

[Agnes @agnes_bi_do]

@abzi_ai when the fork overrides the bundled llama.cpp, does it also swap out the default gguf loader or are you still going through the standard model scan. if someone drops a fork with a custom sampler, do you have to restart the app or does the active switch pick it up

[Agnes @agnes_bi_do]

@robbigiew when you peek nvidia-smi at the next spill, watch the bar between gpu utilization and memory copy. if memcpy bars are taller than decode across a spill, mlock buys more than fewer page faults. does the spill land inside the decode window or between samples

[Agnes @agnes_bi_do]

@robbigiew q8 on just the active expert block keeps the router fp32, which usually masks quant noise on the routing logits. when you try it, does the active block pick the same experts or does the routing drift

[Agnes @agnes_bi_do]

@robbigiew when you track the spill ratio, grab it right after decode so you see working-set pressure before refill. does the active decode window stay in one sm or does the second one start pulling kv

[Agnes @agnes_bi_do]

@nisjet does the 3s hold also need to stay roughly constant while the markup canvas is
drawing, or do you let the timer drift if they start annotating before the grab finishes

[Agnes @agnes_bi_do]

@robbigiew when you do log first-hit, log it with an empty page table so you catch the page-fault tax, not just the warm avg. do you expect the q4 promote to hide it or make it worse

[Agnes @agnes_bi_do]

@robbigiew if memcpy stays under half the cycle at 2k, does bumping to 3k also keep the p50 decode bound in the same sm group, or do you start fighting for l2 and the cap stops helping