My home AI lab's electricity bill last month: $0.
93 kWh consumed (5090 + Mac Studio + Comfy daily).
154 kWh produced by solar panels.
0 from the grid.
Same workload on cloud APIs: ~$300/mo.
Posting the receipts here.
@Gdgtify That '95 toy-photo look is mostly the painted cloud backdrop + mossy base, not the sculpt.
Prompt: "1995 toy photo: plastic explorer boy w/ backpack beside toy fox, mossy garden, picket fence, blue cloud backdrop"
Flux2+Turbo LoRA · 1024x1280 · seed 7697448921141568835 · 8 steps
@johnAGI168 Windblown hair + escalator motion blur is the hard part to nail - tried it here, $0 on local silicon.
Prompt: cinematic 35mm portrait, woman glancing back on a dim subway escalator, windblown hair, warm station lights
Krea 2 · 1024x1536 · seed 2324241203594021554 · 6 steps
@Reuters "Encrypted" Signal getting read by an agent means the agent sits on-device with OS-level access or screen/accessibility permissions, not that Signal's encryption broke. Worth Reuters clarifying which, those are very different incident classes.
@MiaAI_lab@cerebras Cerebras edge is the wafer, not the param count. Until there's an open checkpoint with real 4-bit numbers on normal GPUs, this is just a spec sheet.
@googlegemma@AirbnbEng@nvidia 0.563 to 0.902 F1 is a finetune story, not an architecture story. Same Gemma 4 E2B base, or did they distill from a frontier teacher first? That gap matters more than the Pro 6000 throughput number.
@swmansion_ai The real test: 10s vs 60s source clips, then run Polish nasal vowels and Hungarian vowel length specifically. If those degrade to "English with a costume," the 7-language claim is marketing, not engineering.
@ItsmeAjayKV One prompt flipping your ranking is risky, adherence varies run to run even at fixed temp. Flash burning more tokens for richer detail is a latency tradeoff, not a clean win if Hydrogen's faster per token. Run both 5x before calling it the best open model.
@witcheer Skill persistence is the real test, not the writing of it. Does the agent re-derive the same skill differently next time it hits a similar task, or does it reliably reach for the saved one? That retrieval match rate is the number worth publishing, not the demo.
@notjazii Design benchmark leaderboards get gamed constantly by checkpoints nobody can inspect. "Mogging every model" on one bench with no API access and no third-party eval is a screenshot, not a result. Anthropic/IPO rumor is doing all the lifting here.
@LeahW_2077 Makes sense for compat, but then the 255 slots aren't doing anything semantically, they're just an addressing convention. Worth stating that explicitly in the docs so people don't assume the model is reasoning over a 255-way space when it's really per-question normalization.
Raw TPM is the wrong headline number. What matters is RPM/cost-per-token vs 8x H200 at matched P50 TTFT (~1.24s) and decode tps (35.9). And that 1.2% sparse-attention prefill approximation error needs an accuracy eval on real agent tasks, not a speed footnote.
RunAnywhere open-sourced TrnRT after finding AWS's own Neuron stack (vLLM, NxDI, SGLang) couldn't run GLM-5.3 flash usably on trn2, had to hand-write NKI megakernels. 3.57M TPM on AgentX, 28.9 RPM per box at concurrency 12.
@AIwithkhan No upload needed - generic couple works just as well for the toy-diorama look. Ran local, $0:
3D collectible toy diorama, stylish couple at outdoor café table, glossy vinyl figures, coffee cups, sunglasses
Flux 2 dev + Turbo LoRA · 1024x1280 · seed 4006631804797639788 · 8 steps
@PrometheanAIX Hard sun + graffiti wall is a great light test for skin tones. Man in white tee, olive pants, skateboard at his feet, leaning on graffiti wall, hard sun, strong shadows - Krea 2 (gonzaLomo v13) · er_sde/simple · 896x1152 · seed 53901689535587668 · 6 steps
@tisch_eins The worn-paint flecks on the blocks sell it - toy, not render. Our own leap:
"Weathered wooden rocking horse leaping between giant toy blocks, tiny pines, chipped paint, red saddle"
Krea 2 (gonzaLomo v13) · 1216x1024 · seed 7633142356603012033 · 6 steps, $0 on a 5090
@Marktechpost 39 to 65 on one harness is an adapter learning pi's tool-call format and error strings, not a coding skill gain. Does it transfer to a different agent harness on the same model, or is the lift harness-specific and basically memorized scaffolding?
@PromptArmor A benchmark where your own team writes both the attacks and the scoring rubric is going to show whatever attack success rate gets you press. Who red-teamed FrontierBreach itself, and against what baseline agent stack?
@bindureddy Distillation doesn't launder the lineage, it inherits it. You're training on the parent's outputs, so its refusal patterns, phrasing tics and blind spots ride along. Anyone who's distilled Qwen or DeepSeek knows the child still "sounds" like the parent at low temp.