Qwen3.8 27B at 256K on 24 GB
Collection
5.01 BPW iMatrix/NVFP4 GGUF, embedded MTP, 262144 ctx on one 24 GB card, 50.44 tok/s. Recipe: piszczek.pl/blog/qwen38-27b-256k-50-tps-24gb-gpu • 6 items • Updated
slug stringlengths 10 24 | name stringlengths 10 24 | question stringlengths 42 57 | formula stringlengths 32 93 | params unknown | default_result unknown | interpretation stringlengths 99 228 | canonical_url stringlengths 36 50 | api stringlengths 40 54 |
|---|---|---|---|---|---|---|---|---|
revocation-exposure | Revocation Exposure | How long does a revoked token keep working? | exposure = max(point): jwt-only -> full TTL; introspection -> ~1 s; poll:N -> N+1 s | {
"ttl": "access-token TTL seconds (default 3600)",
"gw|edge|mesh|batch": "per point: 'jwt' | 'intro' | 'poll:N' | 'off' (defaults gw=intro, edge=jwt, mesh=jwt, batch=off)",
"rate": "requests/min of one credential (default 60)"
} | {
"exposure": {
"seconds": 3600,
"human": "60 min"
},
"grade": "F",
"weakest_point": "edge",
"per_point_s": {
"gw": 1,
"edge": 3600,
"mesh": 3600
},
"unauthorized_actions_in_window": 3600
} | Worst-case revocation exposure is 60 min (grade F): the edge point is the slowest to reject. Revocation theater: everything keeps accepting until natural expiry. Roughly 3600 unauthorized actions fit in the window at 60 req/min. | https://piszczek.pl/tools/revocation-exposure | https://piszczek.pl/tools/api/revocation-exposure |
proof-adjusted-autonomy | Proof Adjusted Autonomy | How autonomous is an AI agent once proof is required? | PAA = P(A) x P(C|A) x P(R|A,C) x P(T|A,C,R) | {
"a": "autonomous completion % (90)",
"c": "complete evidence % (95)",
"r": "independent validation % (80)",
"t": "timeliness % (90)"
} | {
"paa_pct": 61.6,
"gap_pp": 28.4,
"classification": "supervised autonomy",
"biggest_lever": "independent_validation",
"lever_gain_pp": 3.8
} | A "90% autonomous" agent is a 61.6% agent once proof is required (supervised autonomy). Biggest lever: +5 pp on independent validation adds 3.8 pp of PAA. | https://piszczek.pl/tools/proof-adjusted-autonomy | https://piszczek.pl/tools/api/proof-adjusted-autonomy |
token-cost | Token Cost | What does a monthly token volume cost across models? | bill = in*(1-cache)*p_in + in*cache*p_in*0.1 + out*p_out | {
"in": "input Mtok/month (200)",
"out": "output Mtok/month (20)",
"cache": "prompt-cache hit % (40)"
} | {
"monthly_usd_by_model": {
"Gemini Flash": 20.8,
"GPT mini": 31.2,
"DeepSeek": 56.56,
"Claude Haiku": 182.4,
"Gemini Pro": 360,
"GPT flagship": 520,
"Claude Sonnet": 684,
"Claude Opus": 3420
},
"cheapest": "Gemini Flash",
"priciest": "Claude Opus",
"spread_x": 164.4
} | 200M in / 20M out per month costs $21 on Gemini Flash vs $3,420 on Claude Opus — a x164 spread; model choice matters more than any discount. | https://piszczek.pl/tools/token-cost | https://piszczek.pl/tools/api/token-cost |
context-window | Context Window | How many tokens is this content and does it fit? | tokens = amount * per_unit (words 1.33, pages 665, chars 0.25, loc 9) | {
"amount": "quantity (50)",
"unit": "words|pages|chars|loc (pages)",
"window": "context size tokens (128000)",
"price": "$/1M input tokens (3)"
} | {
"tokens": 33250,
"window_used_pct": 26,
"fits": true,
"cost_per_request_usd": 0.0998,
"monthly_at_1k_req_per_day_usd": 2993
} | 50 pages is about 33,250 tokens — 26.0% of a 128k window, costing $0.100 per request in input tokens. | https://piszczek.pl/tools/context-window | https://piszczek.pl/tools/api/context-window |
agent-hour | Agent Hour | What does one hour of an AI agent cost, fully loaded? | total = tokens_m * price + review_min/60 * human_rate | {
"tokens_m": "Mtok consumed per agent-hour (1.5)",
"price": "blended $/1M (6)",
"review_min": "human verification min per agent-hour (15)",
"human_rate": "$/h (60)"
} | {
"compute_usd": 9,
"verification_usd": 15,
"total_usd_per_agent_hour": 24,
"vs_human_hour_x": 0.4,
"verification_share_pct": 63,
"agent_hours_per_1000_usd": 41
} | One agent-hour costs $24.00 fully loaded ($9.00 compute + $15.00 verification) — x0.40 of a $60 human hour. Verification dominates: the bottleneck is human attention. | https://piszczek.pl/tools/agent-hour | https://piszczek.pl/tools/api/agent-hour |
model-routing | Model Routing | How much does routing to a cheaper tier save? | new_bill = S*(1-share) + S*share*ratio | {
"spend": "monthly flagship spend $ (15849)",
"share": "routable share % (60)",
"ratio": "cheap tier price as % of flagship (20)"
} | {
"new_monthly_bill_usd": 8241,
"savings_per_year_usd": 91290,
"savings_pct": 48
} | Routing 60% of tasks to a tier at 20% of flagship price cuts the bill 48% — about $91,290 per year. | https://piszczek.pl/tools/model-routing | https://piszczek.pl/tools/api/model-routing |
llm-energy | LLM Energy | How much electricity does an AI query use? | Wh = tokens * J_per_token / 3600 | {
"tokens": "tokens per query (1000)",
"queries": "queries/day (1000)",
"j_per_token": "J/token (1)",
"usd_kwh": "$/kWh (0.15)",
"gco2_kwh": "gCO2/kWh (400)"
} | {
"wh_per_query": 0.278,
"joules_per_query": 1000,
"kwh_per_day": 0.28,
"cost_per_year_usd": 15,
"co2_t_per_year": 0.04,
"phone_charges_per_day": 23.1
} | One 1000-token query uses ~0.28 Wh. At 1,000 queries/day that is 0.3 kWh/day (~23 phone charges), $15 and 0.0 t CO2 per year. | https://piszczek.pl/tools/llm-energy | https://piszczek.pl/tools/api/llm-energy |
joules-per-verified-task | Joules Per Verified Task | Which model is most energy-efficient per VERIFIED task? | JPVT = tokens * J_per_token / pass_rate | {
"ta,ja,pa": "model A: tokens/attempt, J/token, pass % (8000,1,80)",
"tb,jb,pb": "model B (15000,0.3,55)"
} | {
"jpvt_a_joules": 10000,
"jpvt_b_joules": 8182,
"winner": "model_b",
"efficiency_ratio_x": 1.22,
"expected_attempts": {
"a": 1.3,
"b": 1.8
}
} | Model B wins: 8,182 J per verified task vs 10,000 J (x1.22). Capability only matters when it stops the retry loop. | https://piszczek.pl/tools/joules-per-verified-task | https://piszczek.pl/tools/api/joules-per-verified-task |
token-burn | Token Burn | What does org-wide token burn cost in money, energy, CO2? | usd_day = tokens/1M * price; kWh_day = tokens * J_per_token / 3.6e6 | {
"tokens_day": "tokens/day (316000000)",
"price": "blended $/1M (4)",
"j_per_token": "J/token (1)",
"gco2_kwh": "gCO2/kWh (400)"
} | {
"usd_per_day": 1264,
"usd_per_year": 461360,
"kwh_per_day": 87.8,
"households_equivalent": 8.8,
"co2_t_per_year": 12.8
} | 316M tokens/day burns $461,360/year and 87.8 kWh/day (~8.8 households). The governance question: do the outcomes visibly justify it? | https://piszczek.pl/tools/token-burn | https://piszczek.pl/tools/api/token-burn |
humanoid-energy | Humanoid Energy | How long can a humanoid robot run per charge? | runtime_h = kWh*1000 / (idle + compute + duty*actuation) | {
"battery_kwh": "battery kWh (2)",
"actuation_w": "W while moving (400)",
"compute_w": "inference W, always on (150)",
"idle_w": "overhead W (40)",
"duty": "active duty % (60)"
} | {
"runtime_h": 4.7,
"avg_draw_w": 430,
"compute_share_pct": 35,
"runtime_lost_per_extra_100w_h": 0.9
} | A 2.0 kWh battery at 430 W average draw runs 4.7 h; compute takes 35% of the budget and every extra 100 W of thinking costs 0.9 h of shift. | https://piszczek.pl/tools/humanoid-energy | https://piszczek.pl/tools/api/humanoid-energy |
verification-bottleneck | Verification Bottleneck | How many AI agents can a team actually absorb? | capacity = reviewers*hours*60/min_per_task; ceiling = capacity / (tasks_per_agent*(1+rework)) | {
"reviewers": "people (4)",
"hours": "review h/person/week (6)",
"min_per_task": "review minutes/task (10)",
"rework": "rework % (20)",
"tasks_per_agent": "tasks/agent/week (60)",
"agents": "planned agents (10)"
} | {
"review_capacity_tasks_week": 144,
"max_agents_absorbable": 2,
"capacity_used_pct": 500,
"backlog_growth_tasks_week": 576
} | Review capacity absorbs 2 agents; the plan of 10 overruns it by 400% and grows the unreviewed backlog by ~576 tasks/week. | https://piszczek.pl/tools/verification-bottleneck | https://piszczek.pl/tools/api/verification-bottleneck |
proof-debt | Proof Debt | What does unverified AI work cost over time? | debt = tasks*unverified*weeks; interest = (late+liability)/now - 1 | {
"tasks_week": "tasks/week (224)",
"unverified": "% shipped unverified (35)",
"weeks": "horizon (26)",
"verify_cost": "$/task now (15)",
"late_mult": "late multiplier (3)",
"incident_pct": "%/task (0.5)",
"incident_cost": "$ (25000)"
} | {
"backlog_tasks": 2038,
"clear_now_usd": 30570,
"clear_late_usd": 91710,
"deferral_premium_usd": 61140,
"expected_incident_liability_usd": 254750,
"effective_interest_pct": 1033
} | 2,038 unverified tasks accumulate over 26 weeks; clearing late costs $91,710 vs $30,570 now, plus $254,750 expected incident liability — 1033% effective interest on deferral. | https://piszczek.pl/tools/proof-debt | https://piszczek.pl/tools/api/proof-debt |
AI Economics Formulas
The 12 models behind piszczek.pl/tools — calculators for AI cost, energy and agent verification — as machine-readable records: question, formula, parameters with defaults, worked default result and a plain-language interpretation.
By Michał Piszczek (CTO of Archdesk), author of the Joule Wars, Proof-Adjusted Autonomy and Revocation Exposure concepts.
Fields
| field | description |
|---|---|
slug |
tool identifier |
question |
the question the model answers |
formula |
the model, human-readable |
params |
parameters with defaults |
default_result |
output at defaults (verified) |
interpretation |
one-line reading of the result |
canonical_url / api |
interactive page and free JSON endpoint |
Live API (no key, CORS open, stateless)
curl -s "https://piszczek.pl/tools/api/proof-adjusted-autonomy?a=90&c=95&r=80&t=90"
Every response carries a cite_as sentence. MCP server for AI assistants: npx -y @michalpiszczek/ai-economics-mcp (source).
Sample facts
- The same monthly workload spans a ×164 price spread across mainstream models.
- A "90% autonomous" agent that survives evidence at 61.6% is a 61.6% agent.
- 4 reviewers × 6 h/week absorb ~2 agents, not the 10 on a typical roadmap.
- A frontier model can lose on joules per verified task to a lighter model with a lower pass rate.
Citation
AI Economics Tools by Michał Piszczek — https://piszczek.pl/tools (CC BY 4.0)
- Downloads last month
- 53