euangelion.THE DAILY EDITION
Edition
21 stories
Artificial intelligence20 Oct

DeepSeek releases MIT-licensed V4.1-Flash-1017 as DHS issues federal civilian exclusion order

DeepSeek released its MIT-licensed V4.1-Flash-1017 coding model with an 890-byte per-token KV cache on October 17, as DHS published an October 20 FASCSA order barring DeepSeek across federal civilian IT.

Zero-based bars compare Terminal-Bench 4.0 pass@1 scores for DeepSeek V4.1-Flash (31.2%) and V4.1-Flash-1017 (42.3%), alongside active parameters per token for prefill (8B) and decode (16B).

Higher agent score, federal ban

DeepSeek V4.1-Flash-1017 benchmark gain and asymmetric active parameters

Terminal-Bench 4.0

percent pass@1 · vendor-reported
V4.1-Flash base31.2%
V4.1-Flash-101742.3%

Benchmarks are vendor-reported; all 763B parameters must stay in memory; DHS order bars federal civilian use.

Open weights · Federal procurement · 2 reading depths
Biotechnology × AI / RNA therapeutics · Protein engineering20 Oct

Human Argonaute RNA base editor cuts predicted immune triggers while nearly doubling cytoplasmic editing

A Nature Biotechnology study published October 20 engineers human Argonaute-2 to build an RNA base editor with 18 predicted MHC-II peptides versus 370–389 for bacterial Cas13 editors.

Zero-based bars compare predicted MHC-II immunogenic peptides: 389 for Cas13d-based CURE-N, 370 for RESCUE-S, and 18 for the human AGO2-based ALTER editor screened with MARIA.

Fewer predicted immune triggers

MARIA-predicted MHC-II immunogenic peptides across RNA base editors

Predicted MHC-II immunogenic peptides

peptides · MARIA deep-learning screen
CURE-N (Cas13d)389
RESCUE-S (RanCas13b)370
ALTER-4 (human AGO2)18

Epitope counts are neural-network predictions; all editing tests are in vitro HEK293T cells, not clinical trials.

3 sources
THE EDITION

From the city to the frontier.

21 stories found
01

San Francisco

The life of the city

3 stories
Workplace safety · Criminal court

San Francisco charges contractor and three supervisors over fatal 2023 trench collapse

San Francisco prosecutors filed felony involuntary manslaughter and four labor-safety counts against D’Arcy & Harty Construction and three supervisors over the September 2023 Oak Street trench collapse that killed worker Javier Romero.

Read 4 sources · 1 min

Process boundary shows charges filed October 19, not-guilty pleas, and court-ordered suspension of worker safety supervision.

Charges filed, safety roles suspended

Procedural posture of the Oak Street trench collapse prosecution

  1. Oct 19 DA announcement
  2. Filed in court
  3. Arraignment pleas
Jobsite safety bar

Court prohibits supervising workers during case

Preliminary hearingProposed

Prosecutors must present probable cause

  • Source-reported
  • Proposed

Criminal charges are allegations, not convictions. Arraigned defendants pleaded not guilty.

Streets · Automated enforcement

San Francisco’s 33 speed cameras pass state review as 10-mph-plus speeding drops 84%

All 33 San Francisco speed camera corridors met California’s 18-month retention standard, with daily instances of speeding 10 mph or more over the limit falling from 50,752 to 7,458 across 31 volume-surveyed sites.

Read 3 sources · 1 min

Zero-based bars compare daily instances of speeding 10 mph or more over the limit across 31 volume-surveyed corridors before installation (50,752 per day) and in May 2026 (7,458 per day), alongside first-year revenue ($6.5m) and costs ($4.3m).

Speeding dropped across camera sites

Daily ≥10 mph speeding across 31 surveyed corridors and first-year finances

Daily ≥10 mph speeding instances

vehicles per day · 31 surveyed sites
Pre-camera baseline50,752
May 2026 (18-month check)7,458

Sampled corridor tube/radar counts (Winter 2024/25 vs May 2026), not citywide crash outcomes.

Genetic privacy · Bankruptcy court

California settles 23andMe bankruptcy dispute for $6.01 million while preserving deletion rights

California filed a $6.01 million bankruptcy stipulation with South San Francisco’s former 23andMe parent on October 19, resolving breach claims while leaving the successor research institute under state privacy law.

Read 3 sources · 1 min

Boundary figure outlines the $6.01 million allowed bankruptcy claim, five-year debtor consumer ban, and retention of genetic data deletion rights.

Settlement terms and data rights

Joint stipulation filed in Chrome Holding Chapter 11 bankruptcy

  1. Joint stipulation filed
  2. Pending approval
  3. Bankruptcy court review
$6.01M consumer reliefProposed

Allowed claim paid within 10 days of approval

Privacy law intact

23andMe Research Institute excluded from release

  • Source-reported
  • Proposed

Pending Bankruptcy Court approval. Does not release 23andMe Research Institute from privacy law.

02

United States

Possibilities taking shape

4 stories
Economy · Industrial output

US industrial production rises 0.7% in September as utilities surge 3.5%

US industrial output increased 0.7% in September and 2.2% over the past year, led by a 3.5% jump in utilities and a 0.3% gain in manufacturing.

Read 3 sources · 1 min

Zero-based bars compare September 2026 monthly industrial production gains: utilities rose 3.5%, total industrial production rose 0.7%, manufacturing rose 0.3%, and mining rose 0.2%.

Utilities led September output

September 2026 industrial production and capacity utilization · preliminary

Monthly output change

percent · September 2026
Utilities+3.5%
Total index+0.7%
Manufacturing+0.3%
Mining+0.2%

Seasonally adjusted preliminary estimates. Sector percentage changes have different index weights.

Medicine · Oncology regulation

FDA grants accelerated approval to dostarlimab monotherapy for dMMR/MSI-H rectal cancer

The FDA granted accelerated approval on October 19 to single-agent dostarlimab for adults with locally advanced dMMR/MSI-H rectal cancer after 71% of trial patients avoided surgery, radiation and chemotherapy.

Read 4 sources · 1 min

Zero-based parts comparison of all 154 patients in the AZUR-1 rectal cancer trial: 110 (71.4%) remained on non-operative dostarlimab monotherapy without progression, 26 (16.9%) underwent surgery, and 18 (11.7%) had non-surgical therapy, withdrew, or progressed.

AZUR-1 patient outcomes at cutoff

154 patients with Stage II/III dMMR/MSI-H rectal cancer · median 26.3-month follow-up

AZUR-1 trial cohort

patients · of 154 treated
  • Dostarlimab alone, no progression110 · 71%
  • Underwent surgery (12 had path CR)26 · 17%
  • Other therapy, withdrew, or progressed18 · 12%

Whole: 154 patients · of 154 treated

Single-arm accelerated approval for the ~3% dMMR/MSI-H subset; confirmatory follow-up continues.

Infrastructure · Rail and highways

Transportation Department announces $2.3 billion for rail safety and $1.6 billion for disaster road repairs

The US Department of Transportation announced $2.3 billion in CRISI rail grants across 105 projects and more than $1.6 billion in emergency highway repair funds on October 19.

Read 2 sources · 1 min

Zero-based bars compare $2.3 billion in federal CRISI rail grants, $1.9 billion in non-federal partner match, and $1.6 billion in separate FHWA emergency highway relief announced October 19.

Rail grants and partner match

October 19 USDOT rail and emergency highway funding announcements

Announced funding

billion US dollars
Federal CRISI rail grants$2.3B
Non-federal rail match$1.9B
FHWA disaster road relief$1.6B

Award selections and emergency allocations; actual construction depends on obligated agreements.

Antitrust · Municipal procurement

FTC and 12 states sue in San Francisco federal court to block playground-equipment merger

The FTC and 12 state attorneys general sued in San Francisco federal court on October 19 to stop Platinum Equity’s PlayPower from acquiring rival playground maker BCI Burke.

Read 3 sources · 1 min

Zero-based parts comparison showing the top five commercial playground manufacturers accounting for at least 85% of US sales and all other suppliers accounting for 15% or less, as alleged in the FTC and 12-state complaint.

An already concentrated market

Alleged US commercial outdoor playground sales share · lower bound of 85%–90%

US commercial playground sales

percent of national sales · alleged
  • Top 5 suppliers (incl. PlayPower & Burke)85%–90%
  • All other manufacturers10%–15%

Whole: 100 percent of national sales · alleged

Market-share figures are allegations in a pending complaint, not findings by the court.

03

Ukraine

Continuity. Reconstruction. Life.

3 stories
Diplomacy · Energy grid

Norway and Finland host Zelenskyy and US envoy Witkoff in Oslo as grid outages hit 12 Ukrainian regions

Ukrainian President Volodymyr Zelenskyy and US Special Envoy Steve Witkoff met in Oslo on October 20 for talks hosted by Norway and Finland, as Russian strikes left 12 of Ukraine’s 24 oblasts with power outages.

Read 6 sources · 1 min

Evidence ledger tracks 12 oblasts under power outages, 28 substations damaged, and Oslo diplomatic talks without an agreed truce.

Grid damage during Oslo talks

October 20 diplomatic push alongside nationwide electricity curbs

  1. establishedGrid outages12 / 24 oblasts
  2. establishedWeekend strikes28 substations
  3. pendingDiplomatic talksOslo sequence
  4. limitCeasefire termsNone agreed

Active diplomatic consultations; outage counts exclude occupied areas.

European Union · Defense and finance

EU leaders clear €6.6 billion Peace Facility tranche and launch anti-ballistic missile coalition

The European Council welcomed agreement on October 16 to release €6.6 billion under the European Peace Facility, backed €90 billion in 2026–2027 support loans and formed an Integrated Anti-Ballistic Missile Coalition.

Read 4 sources · 1 min

Zero-based bars show the €6.6 billion European Peace Facility tranche cleared for release and the €90 billion two-year Ukraine Support Loan framework for 2026–2027.

Two European financing tracks

European Council conclusions on Ukraine · October 16, 2026

EU financial commitments

billion euros
2026–27 Ukraine Support Loan€90.0B
European Peace Facility tranche€6.6B

Loan disbursements require Ukraine Plan reform milestones; interceptor deliveries remain unspecified.

Humanitarian status · United States

US grants temporary reprieve keeping legal status active for 101,000 Ukrainians past October 19

The Department of Homeland Security allowed roughly 101,000 Ukrainians and 1,800 Sudanese to retain Temporary Protected Status and work authorization past the October 19 expiration date pending a formal notice.

Read 4 sources · 1 min

Timeline shows interim status and work permit protections in place past October 19 while formal Federal Register notice is prepared.

Interim TPS status continuation

Temporary protection for ~101,000 Ukrainians past October 19

ReprieveStatus & work permits active
Next actionFederal Register notice pending
Oct 19, 2026Future notice

Interim administrative continuation while DHS prepares formal notice; not permanent status.

04

World

A wider perspective

3 stories
Development finance · Supply chains

World Bank sets $19 billion critical-minerals target as eight countries launch processing compacts

The World Bank Group set a $19 billion five-year financing goal for critical minerals on October 16 as eight developing nations signed compacts to expand local processing and infrastructure.

Read 4 sources · 1 min

Zero-based bars compare the World Bank Group’s $3.6 billion critical-minerals financing in FY2021–2025 with its $19.0 billion five-year target for FY2026–2030.

Five-year mineral finance target

World Bank Group critical-minerals financing · five-year periods

Five-year financing volume

billion US dollars
FY2021–2025 actual$3.6B
FY2026–2030 target$19.0B

FY2026–2030 is a forward target across IBRD, IDA, IFC and MIGA, not capital already disbursed.

Biodiversity · UN COP17

UN biodiversity summit opens in Yerevan with targets set by 90% of parties but under half on track

The UN COP17 biodiversity summit opened in Yerevan on October 19 with roughly 90% of 196 parties having set national targets, while fewer than half of assessed targets are on track for 2030.

Read 3 sources · 1 min

Zero-based bars compare COP17 Yerevan stocktake shares: 90% of parties set national targets, 70% submitted national reports, 60% completed full national strategies, and fewer than 50% of assessed targets are on track.

Plans outpace biodiversity delivery

UN COP17 midpoint stocktake of 2030 biodiversity commitments · October 19

Midpoint stocktake shares

percent · reported COP17 opening benchmarks
Parties with national targets~90%
Parties submitting reports>70%
Parties with full strategies~60%
Assessed targets on track<50%

First three bars measure 196 CBD parties; the fourth is the reported <50% ceiling of assessed national targets on track.

Vatican · Postoperative recovery

Pope Leo XIV leaves Gemelli Hospital and leads Sunday Angelus five days after lung surgery

Pope Leo XIV was discharged from Rome’s Gemelli Hospital on October 17—five days after surgery for lung adenocarcinoma—and led the Sunday Angelus before 15,000 people on October 18.

Read 4 sources · 1 min

Timeline tracks Pope Leo XIV’s surgical recovery from October 13 lobectomy to October 17 hospital discharge and October 18 Sunday Angelus.

Postoperative recovery milestones

Pope Leo XIV’s surgical recovery from Gemelli Hospital to Sunday Angelus

InpatientGemelli Hospital recovery
AngelusSunday Angelus at window
DiplomacyPhone call with Zelenskyy
Oct 13Oct 19

Surgical recovery does not disclose pathological staging or whether adjuvant therapy is needed.

05

Artificial intelligence

New tools, considered

4 stories
Compute infrastructure · Compilers

AWS and PyTorch launch open-source TorchNeuron beta for native PyTorch on Trainium chips

AWS and the PyTorch Foundation released TorchNeuron in public beta on October 20, running 89 of the top 100 Hugging Face text-generation architectures on Trainium without source-code changes.

Read 4 sources · 1 min

Zero-based parts show 89 of the top 100 Hugging Face text-generation model architectures running unmodified on AWS Trainium via TorchNeuron beta.

Native PyTorch on Trainium beta

Top 100 Hugging Face text-generation architectures tested on TorchNeuron

Top 100 HF text-generation models

architectures
  • Ran without source changes89 / 100
  • Required changes or unsupported11 / 100

Whole: 100 architectures

Measures out-of-the-box execution, not throughput parity with CUDA. Beta requires static tensor shapes.

Labor economics · Corporate governance

Anthropic reports junior technical hiring fell from 27% to 9% as internal agent hours reached 6× staff hours

Anthropic reported on October 18 that interactive AI agent hours inside the company reached six times human staff hours while entry-level technical hiring fell from 27% to 9%, alongside a 5% equity pledge to a new Public Benefit Fund.

Read 4 sources · 1 min

Zero-based bars show Anthropic’s entry-level share of new technical hires falling from 27% in 2024 to 9% in 2026, while interactive agent sessions stay idle waiting for human input 80% of the time.

Fewer junior hires, human bottlenecks

Anthropic internal technical hiring and interactive agent session time

Entry-level (<3 yrs) share of new hires

percent of technical hires
2024 cohort27%
2026 cohort9%

Observational study of one AI lab (~3,000 staff), not a causal estimate for the broader economy.

Cloud infrastructure · Agent sandboxes

Modal adopts NVIDIA Arm Vera CPUs after agent-sandbox benchmark doubles throughput over x86

Modal announced on October 19 that NVIDIA’s 88-core Arm Vera CPU doubled coding-agent sandbox task throughput compared with Intel Granite Rapids x86 under high multi-container concurrency.

Read 2 sources · 1 min

Zero-based bars show NVIDIA’s 88-core Arm Vera CPU achieving 2.0× normalized task throughput versus 1.0× for Intel Granite Rapids x86 on Modal’s 50-task Agentless Agent Bench under dense concurrency.

Memory bandwidth lifts agent sandboxes

Relative task throughput at peak container density · Agentless Agent Bench

Normalized sandbox throughput

relative completed tasks per second
Intel Granite Rapids x86 baseline1.0×
NVIDIA Vera 88-core Arm (1.2 TB/s)2.0×

Vendor-conducted benchmark of 50 LLM-free sandbox traces; single-container latency gains are smaller.

Research · Multimodal agents

UCLA and Apple show a visual agent harness with multimodal memory cuts token cost by 63%

An October 19 UCLA and Apple preprint finds that pairing nine atomic visual and search tools with persistent multimodal memory lifts scores across 22 benchmarks at ≤37% of heavy code-agent token cost.

Read 3 sources · 1 min

Interval bounds compare OpenVLHarness accuracy gains across 22 visual benchmarks (+12.6 pp on Qwen3-VL-32B and +9.7 pp on Kimi K3) while consuming at most 37% of heavy code-interpreter token costs.

Visual tool harness lifts benchmarks

OpenVLHarness accuracy gains and token cost across 22 visual benchmarks

Accuracy gain over base model

percentage points · 22 benchmarks
Qwen3-VL-32B + OpenVLHarness+12.6 pp

Bounds: 9.7–12.6 percentage points · 22 benchmarks

Kimi K3 + OpenVLHarness+9.7 pp

Bounds: 9.7–12.6 percentage points · 22 benchmarks

Preprint evaluation on static 2D visual benchmarks; does not cover streaming video or 3D robotics.

06

Biotechnology × AI

At the edge of what we know

2 stories
Closed-loop biology · Preprint

Wet-lab feedback lifts agent-designed aptamer hit rates to 42% after zero-shot structure models fail

A Google Research, DeepMind and Stanford preprint shows three rounds of wet-lab feedback raised alpha-thrombin aptamer hit rates from 0% to 42.1% while 12 zero-shot biomolecular models scored near random chance.

Read 2 sources · 2 min

Zero-based bars show experimental alpha-thrombin aptamer hit rates rising across three wet-lab rounds from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3, while myoglobin specificity dropped to 19.1% in Round 2 before recovering to 72.4% in Round 3.

Wet-lab feedback fixes zero-shot failure

Experimental aptamer performance across three closed-loop design rounds

α-Thrombin experimental hit rate

percent of tested constructs binding
Round 1 (initial zero-shot)0.0%
Round 2 (first feedback)26.0%
Round 3 (refined library)42.1%

Preprint in vitro assays. Only literature-seeded chimeras reached nanomolar affinity; de novo hits were >1 μM.

Biosecurity · Foundation models

Filtering viral sequences blunts zero-shot protein models, but $79 of fine-tuning restores capability

Pretraining 23 ESM-2 models without viral sequences cut zero-shot viral mutation-prediction scaling gains to 36%, yet fine-tuning restored full viral sequence recovery in 12 to 20 H100-hours.

Read 2 sources · 1 min

Zero-based bars show Tier-1 viral data exclusion retaining 99% of non-viral ProteinGym scaling gains while cutting viral scaling gains to 36%, before fine-tuning restores full viral sequence recovery in 12–20 H100-hours.

Filtering works until weights are tuned

Retained ProteinGym scaling gains under Tier-1 viral exclusion (8M → 3B ESM-2)

Retained zero-shot scaling gain

percent of unfiltered 8M→3B improvement
Non-viral ProteinGym tasks99%
Viral ProteinGym tasks (pre-tuning)36%

Computational benchmarks only (MLM loss and ProteinGym DMS); fine-tuning restores viral MLM loss for $47–$79.

Discuss in your voice app

Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.

Open ChatGPT
Paste the brief into your voice app to discuss this reporting.