daily notes

Cline’s September 2 postmortem explains a dual-bundle loader, activation fallback, shared state and a gradual rollout inside a marketplace without native release cohorts. Its comparable production telemetry put tasks hitting three consecutive mistakes at 6.34% on the old harness and 0.62% on the new one. Read the implementation and measurement sections: this is a specific failure threshold, not a tenfold improvement in every kind of agent reliability.

Epoch fits the state-of-the-art reasoning-model frontier after September 2024 at 14 ECI points a year, versus 6 for non-reasoning models. The article shows the linear fits and 90% prediction intervals rather than presenting the slope as a causal effect. It is worth the full read because the methodology, sensitivity to frontier construction and uncertainty around the short post-o1 window determine how much weight the apparent acceleration deserves.

source post shown above ↑

Artificial Analysis puts Fable 5.1 at 66 on its Intelligence Index, four points above Fable 5, while max-effort task cost rises 20% to $3.76 despite a 75% cache-read price cut. The thread is useful because it decomposes the bill into output-token use, cache savings and effort settings, then shows overlapping confidence intervals on professional-work benchmarks. Read it as an independent harness-specific measurement, not a universal ranking of intelligence.

updates 2026-09-01 · prior evidence ↗

Artificial Analysisoriginal post ↗

Meta has released Muse Spark 1.3, their fourth Muse Spark model release in five months. Muse Spark 1.3 (max), which is in limited preview for Meta’s partners, scores 62 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5.1 and Claude Opus 5. The variant available now, Muse Spark 1.3 (xhigh), scores 61 and ties with GPT-5.6 Sol (max) and Grok 4.6 (high). Both variants’ gains come primarily from improvements in agentic work and scientific capabilities Muse Spark 1.3 (xhigh) enters the Artificial Analysis Intelligence Index at 61, up 4 points from Muse Spark 1.2 (57, August) and 8 points from Muse Spark 1.1 (53, July). It enters tied with GPT-5.6 Sol (max), Grok 4.6 (high), and Claude Opus 5 (high), and behind Claude Fable 5.1 (max, 66), Claude Opus 5 (max, 63), and Claude Fable 5 (max, 62) Muse Spark 1.3 (max), which is in a limited preview stage, lands at 62. This higher index score is enabled by gains vs. Muse Spark 1.3 (xhigh) in Tau3-Bench Banking (52% vs. 47%) and GDPval-AA v2 (1,754 Elo vs. 1,709). Muse Spark 1.3 (max) is second only to Claude’s Fable and Opus variants in total score Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release! Key Takeaways: ➤ Continued improvement on agentic knowledge work tasks. At the launch of Muse Spark 1.2, we noted its significant gains in agentic knowledge work performance vs. Muse Spark 1.1. The latest iteration continues this trend, with Muse Spark 1.3 (xhigh) demonstrating a notable 12-point gain vs. Muse Spark 1.2 in Tau3-Bench Banking (35% to 47%), a 5-point gain in Terminal-Bench 2.1 (80% to 85%), and a new GDPval-AA v2 Elo of 1709 against its predecessor’s 1615. Muse Spark 1.3 (max) improves further on Tau3-Bench Banking (52%) and GDPval-AA v2 (1,754 Elo). This Tau3-Bench Banking score is #1 among all models. Muse Spark 1.3 (max) achieves these higher agentic work scores by using more turns and total reasoning tokens, reasoning 62% more on GDPval-AA v2 and 28% more on Tau3-Bench Banking compared to Muse Spark 1.3 (xhigh) ➤ The lowest cost per task for any model at 59+ on the Artificial Analysis Intelligence Index. Muse Spark 1.3 (xhigh) costs $0.55 per Intelligence Index task at Meta's unchanged $1.25/$4.25 per 1M token pricing ($0.15 for cached input), with its peers GPT-5.6 Sol (max) and Grok 4.6 (high) costing $0.95 and $0.94 respectively, a 70%+ premium. This places Muse Spark 1.3 (xhigh) on the Pareto frontier for Intelligence vs. Cost per Task. Its cost per task is higher than Muse Spark 1.2 ($0.40 per task), driven by ~57% more input tokens per task on agentic evaluations, with output tokens up only ~8%. Pricing for Muse Spark 1.3 (max) is not yet publicly available ➤ Scientific Reasoning results rose across the board, led by CritPt. CritPt was the standout non-agentic score gain vs. Muse Spark 1.2, with a material +8 points for the xhigh variant (18% to 26%), and GPQA Diamond achieved +4 points (90% to 94%), while Humanity’s Last Exam and SciCode each gained a more modest 2-3 points (45% to 47% and 56% to 59%, respectively). Muse Spark 1.3 (max) achieved roughly similar scores to the xhigh variant, gaining 2 points in Humanity’s Last Exam, tying on GPQA Diamond, and losing a point on CritPt vs. Muse Spark 1.3 (xhigh) ➤ Minor regressions in only two evaluations. Both Muse Spark 1.3 (xhigh) and Muse Spark 1.3 (max) dropped 4 points in AA-LCR (83% to 79%) when compared to Muse Spark 1.2, and AA-Omniscience (Accuracy) fell 3 points for xhigh and 1 point for max. The drops in AA-Omniscience (Accuracy) are due to a higher abstention rate (not answering questions when unsure), which also lowered the hallucination rate for Muse Spark 1.3 (xhigh) Other model details (xhigh variant): ➤ Context window: 1M tokens, unchanged from Muse Spark 1.2 ➤ Pricing: unchanged from Muse Spark 1.2: $1.25/$4.25 per 1M input/output tokens, with cache hits discounted to $0.15 per 1M ➤ Input modalities: text, image, video ➤ Availability: Meta's first-party API and Muse Code

Artificial Analysis’s later September 2 evaluation puts the available Muse Spark 1.3 xhigh at 61 and $0.55 per task, versus Gemini 3.8 Flash high at 59 and $0.58. That supersedes the earlier Flash report’s cheapest-at-this-level claim within this harness. Muse’s max variant scores 62 but remains a limited preview with unpublished pricing; the available xhigh variant also regresses on long-context reasoning.

@asimovincoriginal post ↗

Anybody can build an Asimov 1 now. The latest design update strips steps out of the assembly, opens up the parts you need to reach, and puts metal where plastic used to break. New CAD files are on GitHub. https://t.co/fX6pwoqFod

discussion3 selected replies

@asimovincreply ↗

Every screw is in the CAD. Hover over one and it tells you M4x8. You know the position, the count and the type before you pick up a tool. https://t.co/CsUdW2h7UQ

@asimovincreply ↗

The arms use heat-set inserts. You never place a nut blind again. Miss one before the motor went in and you used to take the whole arm apart. https://t.co/Q9iO3MSDKB

@asimovincreply ↗

Two panels are gone. The small back panel, and the covers over the ankle mechanism. Fewer parts to remove, fewer parts to lose. https://t.co/GMkqACTSTT

Asimov published a new CAD design aimed at simpler assembly, easier access and more durable components. The release makes the robot more inspectable and buildable, but its public thread supplies no independent measurements of assembly time or failure rates.

Shanu Matheworiginal post ↗

big fan and learn a lot from both of these guys but some of these numbers are hard to grasp here if I'm understanding correctly. tbf don't think they're presenting this as a base case, but also don't explicitly label it as the ultra bull case either. tried to do the "what I have to believe" for the rate of change and its eye watering. they're the AI experts, so can weigh appropriately but just humbly doing some math on the comments. Compute >Today: $10B-$15B/GW-year of contracted compute rent. Assuming 500K accelerators/GW and 85% utilization, that is $2.70–$4/GPU-hour. >Takeoff: $25B-$50B/GW-year, or $6.70-$13.40/GPU-hour. Dylan says this is what labs may need to pay to capture 70%-80% of new compute. >The jump: compute pricing rises 2x–4x while we're adding 10s of GW per year (more below). At $50B of peak inference revenue today, the spread over compute is $35B-$40B/GW. At $70B-$80B of future blended revenue density, the spread is $20B-$55B against scarcity pricing applied more broadly across entire fleets. Revenue >Today: Anthropic inference revenue has reached $50B/GW annualized on its highest-monetizing capacity. That is a peak on the inference slice, not its fleet-wide average. >Takeoff: Dylan sees $70B-$80B/GW blended by YE27 and $100B+/GW in the unconstrained case. Those equal $18.80–$21.50 and $26.90+ per utilized GPU-hour. >The jump: At 100 GW, applying $50B-$80B only to a 40% inference share produces $2T-$3.2T of revenue. Applying the $70B–$80B blended figure across the whole fleet produces $7T–$8T. At $3T, that is 1.7x Mag4 TTM revenue, 47% of global IT spending and 9% of US GDP. Physical build >Takeoff: Global AI IT additions: 30 GW in 2026 → 50 GW in 2027 → 70 GW in 2028 → 90-100 GW in 2029. >Implied US takeoff: Dylan has separately put US 2026 additions at ~20 GW of IT load = US near two-thirds of this year deployment held steady gives 20 → 33 → 47 → 60–67 GW. At 1.2 PUE, annual facility-power additions are +24 (2026) → +40 (2027)→ +56 (2028) → +72–80 GW (2029). >The jump: That requires FTM + BTM delivery to more than double by 2028 and triple by 2029. SemiAnalysis is more bullish this year than the mid-teens numbers I've been able to reconcile in 2026 and you still need to 3x from that year over 3 years. Capital >Today: combined cloud RPO/backlog is ~$2.3T. Top-five hyperscaler capex is expected to rise from $800B-$860B in 2026 to $1.1T-$1.3T in 2027. >Takeoff: 70%-80% of 70 GW means 49-56 GW of lab additions in 2028. At $40B-$60B/GW, that is $2T+ of capex (and equipment is moving up - i.e. memory + accelerators). At $25B-$50B/GW to rent compute, it is $1.2T-$2.8T of annual compute rent expense. >The jump: $2T-$2.8T is 1.5x-2.5x the entire top-five hyperscaler capex expected for 2027. Current economics vs. takeoff endpoint probably belong in separate cases but the rate of change btwn is insane regardless. I’m sure I’m missing something, but this is how I read the rate of change math.

Mathew translates a bullish AI-infrastructure takeoff case into compute prices, revenue per gigawatt, physical power additions and capital requirements. The arithmetic makes the burden explicit: a 100 GW fleet at the cited revenue densities reaches trillions of dollars, while a 2028 lab build could require more annual compute rent and capex than current hyperscaler plans. The post is valuable as a ‘what must be true’ bridge, not a forecast; several inputs come from the scenario it is challenging.

Adrian argues that shock-prone monetary policy needs conditional scenarios and a legible reaction function rather than precise rate-path commitments. Forecasts should carry risks and escape clauses, while normal market moves around incoming data can improve price discovery instead of representing a communications failure. The full article is a compact framework for judging central-bank guidance when fiscal, energy and geopolitical conditions move faster than a fixed path.

O'Trakoun separates the headline unemployment rate from the declining probability that long-term unemployed workers find jobs. The article distinguishes 27-52 week spells from unemployment lasting at least 53 weeks and uses the outflow rate to show why a stable aggregate rate can hide worsening reemployment prospects. It is worth reading because duration composition can change the labor-market signal before the headline rate does.

Broadcom reported $16.7B of AI semiconductor revenue in fiscal Q3 ended August 2, up 221% year over year and 54% sequentially. The September 2 release guides fiscal Q4 AI revenue to $21.7B and total revenue to $34.8B. The reported quarter adds direct evidence of custom-accelerator and networking demand; the next-quarter figures are management forecasts.

For the week ended August 28, EIA’s September 2 estimates show a 4.5M-barrel commercial crude draw and a 0.8M-barrel distillate build. Distillate stocks remain 14% below the seasonal five-year average, while four-week products supplied fell 4% year over year versus 3% in the prior release. The build has not closed the product-stock shortfall, and the supplied-volume measure weakens the demand-boom explanation.

updates 2026-08-27 · prior evidence ↗

@NickTimiraosoriginal post ↗

Bessent says he expects interest rates will fall once "we get on the other side" of the Iran conflict, and that today's AI investments could be "extremely disinflationary" with results showing up "in the next six months." "The economy here is very, very strong and I think we're accelerating. And this is in the face of this Iran conflict. Interest rates will come down when we get on the other side of this. The economy will accelerate. And I would add too that right now we have what Alan Greenspan would've called a conundrum, that we have large borrowings by AI institutions that are businesses that are meeting these incredible CapEx needs. At a point, as we heard from our private sector partners yesterday, this CapEx will turn into productivity and that will be extremely disinflationary. So I would guess that in the next six months, we will start seeing the benefits of that."

Treasury Secretary Scott Bessent forecast that AI capex will turn into productivity gains that become ‘extremely disinflationary,’ with benefits starting within six months. The deadline makes the claim testable against productivity, unit-cost and inflation data; it remains a policy forecast, not an observed disinflation result.

@nikkeioriginal post ↗

日経平均、終値1889円安　世界金利高でよぎる「うたげの終わり」 https://t.co/1Dau0th3l0

The Nikkei closed 1,889.70 points lower, down 2.85%, as rising global yields weighed on equities. That extends yesterday’s Japanese 10-year yield milestone into a cross-asset selloff; one session does not establish how long the pressure will last.

updates 2026-09-01 · prior evidence ↗

A new reserve decomposition separates active portfolio shifts from changes in countries’ reserve size. For 2019–23, the 62 countries with complete data contributed little to the aggregate dollar-share decline; the researchers infer a concentrated effect from four large holders with missing data. That changes how the aggregate trend should be read, but the missing allocations prevent a directly observed country-by-country attribution.