The future is arriving badly explained.
In brief.
Updated on · Evening edition. Four signals worth your attention.
ASML sees demand outgrowing its machine-building plans
ASML says its planned supply for 2027 and 2028 will not fully meet customer demand. The chipmaking-equipment supplier is investigating additional capacity for 2028. The bottleneck is the equipment used to make chips, not just the chips themselves; further expansion remains under study, not completed supply.
KDDI reports faster uploads after AI retunes its network
KDDI reports about 60% faster average uploads during congestion in 100 selected areas of its commercial network. Multiple AI systems adjusted control-signal resources as traffic changed during a June–August trial. The carrier plans wider deployment from November; this company-reported result is not a guaranteed gain for every subscriber.
A robotics lab records the forces behind human movement
South Korea’s KITECH unveiled a platform that synchronizes joint motion, foot forces, hand movements and muscle signals. It collects human movement data for humanoid control and compares people’s movements with and without wearable robots under matched conditions. This is measurement infrastructure, not evidence that a humanoid has learned a new skill.
Nine months of New England wind power beat last year’s total
EIA reports that wind farms in the ISO New England grid area generated 4,619 gigawatt-hours from January through September, already exceeding 4,533 in all of 2025. That is electricity produced, not just capacity announced. The comparison covers unequal periods; it is not a full-year forecast or a guarantee of supply.
Sources published on 14 October 2026. KDDI’s trial ran in June–August; EIA’s wind figures compare January–September 2026 with all of 2025. Company and institute claims are attributed, not independently demonstrated deployment outcomes.
In Charts.
Battery Incidents Barely Fell. Capacity Almost Doubled.
These are report-snapshot values, not a complete count of every battery failure. Annual events divided by installed GWh are not an individual site's failure probability. The public CSV has 23 events for 2025, versus 24 in the report; the versions are not combined.
The denominator does most of the moving
For a community considering a large battery installation, a falling incident rate sounds reassuring. But the numerator matters too: did incidents become rare, or did the amount of installed storage grow much faster?
In EPRI's October 2026 report, the recorded annual count falls from 26 to 24, a decline of 7.7%. The published capacity figure rises from 343 to 634 gigawatt-hours, up 84.8%. Divide one by the other and the result falls from 7.58 to 3.79 incidents per 100 GWh, a 50.1% decline. The large change in that ratio is not a halving of the incident count.
A useful ratio is not a site-risk forecast
The report tracks major internal thermal incidents in stationary lithium-ion storage, including utility and commercial or industrial installations. It is not a census of every malfunction, electric-car battery or home battery. Reporting is incomplete, and each incident counts once regardless of its severity.
The denominator is stored-energy capacity, not electricity generated, power in gigawatts or operating years. Newly installed capacity can enter the annual stock without having operated for a full year. We also could not independently verify that the proprietary capacity series covers every installation type in the incident count.
The fair reading is narrower than “batteries are half as dangerous”: in this report, recorded incidents fell slightly while the installed capacity denominator rose sharply between these two years. A decision about one installation still depends on its design, operation and consequences of a failure.
Sources, Method & Data
Snapshot and calculation. Four aggregate values come from printed pages 6–7 of EPRI report 3002034920, dated October 2026. Its PDF was created on October 8; the exact first-publication day was not established. This chart is a new explanation of 2024–2025 data, not a new incident count for October 2026. Each ratio is annual recorded events / cumulative installed capacity in GWh × 100. Changes compare 2025 with 2024. Input capacities are already rounded; two decimal places on the ratio do not imply extra measurement precision.
Source versions do not fully reconcile. The public stationary-storage CSV retrieved on October 10 carried a September 29, 2026 Last-Modified header. It contains 26 events dated 2024 and 23 dated 2025, versus 26 and 24 in the newer report. Some regional and confidential-case totals also differ. A version difference is plausible, but no case-by-case revision history resolves it. We reproduce the report aggregates, not a reconstructed current global total; no cases have been invented or added to the CSV.
Coverage and denominator. EPRI includes major thermal events caused by internal battery-system failures and excludes external causes such as wildfires. Some cases are confidential. The report's source figure cites Benchmark Mineral Intelligence's BESS Forecast Q1 2026, China Energy Storage Alliance and Korea Electrical Safety Corporation. The underlying capacity records were not independently accessible. Matching of utility, commercial/industrial and UPS coverage between numerator and denominator remains unverified. This simple annual quotient is not exposure in GWh-years, an individual failure probability, a severity-adjusted measure or the report's separate age-based analysis.
| Year | Recorded incidents | Capacity (GWh) | Incidents per 100 GWh |
|---|---|---|---|
| 2024 | 26 | 343 | 7.58 |
| 2025 | 24 | 634 | 3.79 |
Graphic and rights. Own calculation and original graphic using four published aggregate facts; no EPRI figure or layout reproduced. EPRI's report is © 2026, all rights reserved, not an open-licensed dataset. The accessible CSV is linked for source-version checking, not substituted for the report.
EPRI's October report records 26 incidents in 2024 and 24 in 2025. A much larger capacity denominator roughly halves the ratio, not an individual battery's risk.
Explore the chartCan AI Customer Service Solve the Same Case Four Times?
Numbers inside the bars count cases, not individual runs. “Every time” means success in all four recorded runs, not four chances to succeed once. Two examples, not a model ranking.
Getting it right once is the easier bar
Imagine asking customer service to replace a lost card. A successful interaction is encouraging. The harder question is whether the same request still works when the conversation takes a different path.
In these published simulations, Claude Opus 5 succeeded at least once on 64 of 97 cases; Qwen 3.8 Max did so on 71. Require success in all four runs, and those counts fall to 31 and 34. The middle segment shows cases that sometimes worked and sometimes did not.
A benchmark, not a bank
These are recorded evaluations in a fictional banking environment, with an AI playing the customer. Vastkind grouped the original scores; we did not run the agents again or independently regrade their conversations. The chart measures neither real customer satisfaction nor financial safety.
Both agents received the same embedded case definitions and tool access, but used different reasoning settings. These two examples do not establish a general model ranking or equal-cost comparison. Four runs per case are a small sample, not a guarantee about the next run.
A further reproducibility limit: 15 of the 97 embedded task definitions differ from the code revision named in the run files. They match between the two files analyzed here. The reason for the repository discrepancy is unresolved.
Sources, Method & Data
What we counted. The τ-knowledge banking_knowledge AllTools run files contain 97 case definitions and 388 runs per model, four distinct trial IDs per case. We treated a logged reward within 0.000001 of 1 as success, matching the project metric, then grouped cases by zero, one to three, or four successes. Both selected files contain no infrastructure-error terminations. Bar widths are group counts divided by 97; labels give the counts.
Settings and timing. Qwen 3.8 Max used xhigh reasoning; Claude Opus 5 used max. Both used GPT-5.2 with low reasoning as the customer simulator and a 200-step limit. The metadata evaluation date is August 3, 2026; raw timestamps span August 3–4 without a stated time zone. The submission date is August 4; the hosted files also carry August 4 modification timestamps. This is a new explanation of that snapshot, not a new October test or a comprehensive current leaderboard.
Denominators matter. Individual-run success was 189/388 for Claude and 214/388 for Qwen. That average is not the first attempt. “At least once” counts cases with one or more successes; “every time” counts cases with four. We make no independence assumption or statistical-significance claim.
Original records. Repository snapshot; MIT code license. Own aggregation and graphic, not a copied source figure. The hosted conversation files carry no separate license notice; full dialogs are not republished here.
- Claude Opus 5: submission metadata; original run file (about 240 MB). SHA-256:
80bd60cd0dc19855e685b85ef2c40efcbc2dabae1ffa16988b40b781cbeee2a3. - Qwen 3.8 Max: submission metadata; original run file (about 240 MB). SHA-256:
8c8191c43dfb2d21c1322cc154740e5e6044151837ab817c2cbbcd13ffeb626e.
Task-version boundary. The run files name code commit fc0055dc4e0a316c3f83133267fbd6faaa770992. Fifteen embedded task definitions differ from that revision and from the repository snapshot above. Our counts describe the scores in the published run files. They do not establish an exact reproduction from the named code revision. Other model results were not included because their comparability was not fully verified.
| Model | 0 of 4 | 1 of 4 | 2 of 4 | 3 of 4 | 4 of 4 | Total |
|---|---|---|---|---|---|---|
| Claude Opus 5 | 33 | 11 | 12 | 10 | 31 | 97 |
| Qwen 3.8 Max | 26 | 14 | 5 | 18 | 34 | 97 |
Related: Can AI spot an IKEA assembly mistake? A different capability test, with its own limits.
Two AI agents, 97 simulated cases each, four runs per case. Most cases succeed at least once. Far fewer succeed every time.
Explore the chartCan AI Spot Your IKEA Mistake?
Lime line: successive best scores among tested models. Gray dots: other tested models. 60 photos, three furniture builds; photos + manual + tools. A retrospective comparison, not a physical assembly test.
A practical test of visual reasoning
A photo, an assembly manual and tools. Epoch’s test asks models to judge the assembly and, if something is wrong, identify the relevant steps and explain the mistake. The published benchmark score rose from 28.3% for Claude Opus 4.5 to 80% for GPT-6 Astra.
Not a promise to fix your furniture
The sample covers only 60 photos from three builds and is partly graded by another AI. It does not establish reliability on other furniture, real-time repairs or physical assembly. The score is not simply the proportion of flawed builds flagged, and the tested models are not a complete historical leaderboard.
Sources, Method & Data
The 21 source rows are ordered by model release date as reported by Epoch, not test date. The line retains each new record among those models; it is not a complete history of all available AI. GPT-5 and Gemini 3 are absent. Published scores are plotted unchanged ×100. The endpoints are 28.3333% and 80%, a gain of 51.6667 percentage points. Source standard errors are 5.8665 and 4.9289 percentage points; all model standard errors are in the table below. They are not 95% confidence intervals.
The dataset has 42 intentionally flawed and 18 correct photos from STÄLL, TONSTAD and GULLABERG builds. Models receive instructions and image/Python tools. Correct step identification and an adequate error description matter; a lenient LLM grades descriptions. The score is not simply the fraction of flawed builds flagged. Some exported scores imply half cases when multiplied by 60; the public method does not explain that granularity, so no exact “x out of 60” count is inferred. No human-performance comparison. Accessed October 6, 2026.
Benchmark methodology · Original chart dataset
| Model | Release date (Epoch) | Score (%) | Standard error (pp) |
|---|---|---|---|
| Claude Opus 4.5 | 2025-11-24 | 28.3333 | 5.8665 |
| GPT-5.2 | 2025-12-11 | 38.3333 | 6.3298 |
| Claude Opus 4.6 | 2026-02-05 | 28.3333 | 5.8665 |
| Gemini 3.1 Pro | 2026-02-19 | 26.6667 | 5.6332 |
| GPT-5.4 | 2026-03-05 | 37.5000 | 6.2465 |
| Claude Opus 4.7 | 2026-04-16 | 33.3333 | 6.1372 |
| Kimi K2.6 | 2026-04-20 | 21.6667 | 5.2301 |
| GPT-5.5 | 2026-04-23 | 44.1667 | 6.4101 |
| Claude Opus 4.8 | 2026-05-28 | 42.5000 | 6.3807 |
| Claude Fable 5 | 2026-06-09 | 35.8333 | 6.1859 |
| GPT-5.6 Sol | 2026-07-09 | 56.6667 | 6.3409 |
| GPT-5.6 Terra | 2026-07-09 | 54.1667 | 6.4321 |
| GPT-5.6 Luna | 2026-07-09 | 42.5000 | 6.3807 |
| Kimi K3 | 2026-07-16 | 34.1667 | 6.1170 |
| Gemini 3.6 Flash | 2026-07-21 | 23.3333 | 5.3766 |
| Claude Opus 5 | 2026-07-24 | 60.8333 | 6.1859 |
| Gemini 3.7 Flash | 2026-08-13 | 26.6667 | 5.6332 |
| Claude Fable 5.1 | 2026-09-01 | 70.0000 | 5.7244 |
| Qwen3.8 Max | 2026-09-01 | 20.0000 | 5.2076 |
| Gemini 3.8 Flash | 2026-09-02 | 31.6667 | 5.9383 |
| GPT-6 Astra | 2026-09-03 | 80.0000 | 4.9289 |
A photo, an assembly manual and tools. The best score among tested models rose from 28% to 80% in Epoch’s furniture test.
Explore the chartLatest
Pick a frontier.
Seven fields. A clear place to begin.
Stay ahead
of the story.
A personal take on the week. By Dorian.
Occasional editions. Free to subscribe.
Unsubscribe anytime.