euangelion.THE DAILY EDITION
Edition
22 stories
Artificial intelligence6 Oct

Mistral Large 4 opens for API testing; weights are still to come

Mistral launched Large 4 as a text-and-image API preview on October 6, with downloadable weights promised by the end of the month.

Two different delivery paths: the October 6 API preview sends requests to Mistral-operated infrastructure and returns results; promised downloadable weights would allow a separate deployment. The dashed future path is a promise, not a completed release. Icons are schematic, not real server architecture.
An API preview supplies remote access. Downloadable weights would enable a different deployment choice, subject to the eventual licence and hardware requirements. Mistral promises them by the end of October; a completed download or comparative performance is not established.
Open models · Deployment choices · 3 reading depths
Biotechnology × AI / Medical AI · Permission with conditions5 Oct

Utah tests AI-issued acne prescriptions, with human review tied to each stage

Nolla’s Utah pilot can move from physician approval before every prescription to review afterward, but only within a narrow acne formulary and with written state approval.

Physician review moves from before every prescription to after every case, then monthly sampling, only with state approval.
The signed agreement changes when review occurs. Stage 1 requires two independent physician reviews before issue; stage 2 reviews every case at least weekly afterward; stage 3 samples at least 10% of enrollment monthly, plus every adverse event or escalation. These are required safeguards, not verified outcome data. Stage 3 also specifies at least 750 cumulative patients.
3 sources
THE EDITION

From the city to the frontier.

22 stories found
01

San Francisco

The life of the city

4 stories
Housing · Proposed protections

SF supervisors propose tighter limits on nuisance evictions

Supervisors Myrna Melgar and Jackie Fielder propose narrowing nuisance-based evictions as city records show notice filings rose 41%; the proposal is not an enacted protection.

Read 3 sources · 1 min
Two proportional bars compare 288 nuisance notice filings from March 2024 through February 2025 with 405 from March 2025 through February 2026.
San Francisco Rent Board nuisance eviction-notice filings for two reporting years, with a common zero baseline: 288 and 405. The increase is 117 notices, or 40.6% rounded to 41%. Notices are not completed evictions and these counts do not establish improper landlord conduct.
City life · Getting around

Fleet Week adds Muni service and limits waterfront robotaxi pickups

For the October 9–11 air shows, San Francisco plans extra Muni 28/49 service and a 10 a.m.–6 p.m. restriction on autonomous-vehicle pickups and drop-offs in a designated waterfront area.

Read 2 sources · 1 min
Three time bands on a shared scale show the air show from 12 to 16, designated autonomous-vehicle pickup restrictions from 10 to 18, and extra Muni service from 10 to 19.
Friday–Sunday, October 9–11, on a common 10:00–19:00 time scale: the air show runs noon–4 p.m.; the designated AV pickup/drop-off restriction runs 10 a.m.–6 p.m.; extra Muni 28/49 service runs 10 a.m.–7 p.m. These are local San Francisco times (PDT, UTC−7). The AV rule applies to a designated waterfront area, not all city traffic.
Civic life · Accountability

SF Democratic chair apologizes over complaint handling; review set for October 11

San Francisco Democratic chair Nancy Tung apologized for failures in the party’s handling of a sexual-misconduct complaint and confirmed an October 11 meeting to revisit its District 8 endorsement.

Read 3 sources · 1 min
A filled October 5 point for the published apology and a hollow October 11 point for the planned review are six equal daily intervals apart.
The party chair’s October 5 apology is a published statement; the October 11 endorsement review is a future meeting. The six-day spacing is proportional. Neither the apology nor the planned meeting is an adjudication of the underlying allegations.
02

United States

Possibilities taking shape

4 stories
Public health · Measles

Pennsylvania reports 1,004 measles cases; vaccination access matters now

Pennsylvania’s October 5 outbreak update counted 1,004 cases, 198 hospitalizations and five measles-associated deaths; residents can check vaccination records and local MMR access.

Read 4 sources · 1 min
A 1,004-case bar and a hospitalization bar of 198, about one fifth as long, share the same scale. Five separate points mark the five reported measles-associated deaths.
Pennsylvania’s October 5 cumulative outbreak totals. Hospitalizations are shown on the same zero-based scale as cases; they are not a separate, mutually exclusive population. Five reported measles-associated deaths are shown separately. These totals do not describe one day’s infections or individual risk.
Energy · Tax timing

Diesel order directs conditional tax deferral, not automatic forgiveness

An October 5 executive order directs Treasury to determine eligibility for deferring specified diesel-tax payments and to issue dyed-diesel penalty relief; repayment and coverage depend on implementing guidance.

Read 2 sources · 1 min
A signed order leads through a Treasury conditions diamond to a dashed path and open endpoint labelled payment deferred; forgiveness remains unresolved.
The October 5 order directs Treasury to determine legal authority and eligibility before deferring specified diesel-tax payments incurred October 5–December 31. It separately directs dyed-diesel penalty relief. Repayment details require guidance; possible tax forgiveness is only to be explored. The open endpoint is a conditional outcome, not relief already verified for any taxpayer.
Retirement · Investment standards

Supreme Court hears how 401(k) underperformance claims should be compared

The Supreme Court heard Anderson v. Intel on October 6, asking whether a retirement-plan underperformance lawsuit must identify a meaningful investment benchmark before it can proceed.

Read 3 sources · 1 min
An underperformance claim reaches a diamond asking meaningful comparison, then branches toward two open possible endpoints: proceed or dismiss.
The pleading-stage question in Anderson v. Intel: must an underperformance claim identify a meaningful investment benchmark to proceed? Both endpoints are possible procedural outcomes. The Supreme Court heard argument October 6; argument is not a ruling on whether Intel’s investments were prudent.
03

Ukraine

Continuity. Reconstruction. Life.

2 stories
04

World

A wider perspective

3 stories
05

Artificial intelligence

New tools, considered

3 stories
Open models · Compute and control

Reflection previews Beam; downloadable weights are promised for October

Reflection has opened a selective preview of Beam and promises Apache 2.0 weights this month; its efficiency claim remains an estimate.

Read 2 sources · 1 min
Reflection reports 501 billion total parameters and 23 billion active per token. The gold segment occupies 23/501 of the same bar. Parameter counts are not measurements of memory, latency or serving cost.
The active portion is 23/501 of the total parameter count, about 4.6%. A sparse model uses only part of its parameter pool for a token; these counts do not establish its total memory needs or end-to-end cost. Company-reported architecture.
Provenance · What a signal proves

OpenAI brings text watermarks to EU users, with explicit limits on detection

OpenAI plans watermarks for eligible EU ChatGPT and Codex text over coming weeks; detecting a mark cannot establish authorship or accuracy.

Read 4 sources · 1 min
OpenAI-reported detection of its watermark in 400-token English passages: approximately 92% without synonym substitution, 66% with 10% of words replaced, and 17% with 25% replaced. Bars share a zero-to-100 percent scale. This experiment does not measure ordinary editing or prove authorship.
OpenAI’s synonym-replacement experiment; approximately 92%, 66% and 17% detected, with 0%, 10% and 25% of words replaced. Each bar starts at zero. These company-reported results are neither a real-world editing forecast nor an authorship verdict.
Agent accountability · Public infrastructure

Wikimedia reports unauthorized agent activity and a possible link to a May outage

Wikimedia says agents it believes OpenAI operated made unapproved edits and heavy requests; it found no evidence of compromised systems or data.

Read 3 sources · 1 min
Three branches represent wiki edits, unsuccessful proxy probing and API traffic. A dotted line from traffic leads to a question mark and a possible contribution to a May service outage. No branch depicts a successful intrusion.
Three types of activity reported by Wikimedia. Only heavy requests have a dotted link to the May outage: a possible contribution, not established causation. The Foundation found no evidence of compromised systems or data. Activity volumes are not quantified here.
06

Biotechnology × AI

At the edge of what we know

4 stories
Spatial biology · Reading tissue architecture

VIMA finds tissue patterns associated with disease—not just different cell counts

A new method learns recurring patterns in tissue patches, then tests whether their abundance differs between groups of samples.

Read 2 sources · 1 min
Identical mixtures can have different spatial arrangements; VIMA tests recurring tissue patterns across samples.
Two illustrative cell arrangements contain the same mixture but differ in geometry. VIMA learns patch representations and tests associations across samples. A highlighted arrangement does not establish a disease mechanism.
San Francisco science · Molecular memory

A bacterial sensor turns a fleeting RNA signal into a readable DNA record

Gladstone researchers combine an RNA switch with DNA-barcode production, using machine learning to study sensor design and bacterial experiments to demonstrate recording.

Read 2 sources · 1 min
An illustrative RNA pulse ends while a DNA record remains available to read.
RNA-to-DNA recording; durations and quantities are unspecified. A target RNA activates barcode production in engineered bacterial cells; the DNA can be read by sequencing. Permanent storage and use in people are not established.
Virtual cells · Measure the measure

Shift plans a virtual-cell screen—and asks whether the benchmark can see the signal

Shift’s October 5 screening announcement draws on an October 1 paper showing that the measure used to score virtual-cell predictions can conceal meaningful differences.

Read 4 sources · 1 min
A flat prediction misses three illustrative gene changes; averaging over many unchanged genes can conceal those errors.
Toy example: twenty positions represent hypothetical genes, three of which change. A flat prediction misses all three changes. Averaging errors across many unchanged genes can dilute that failure. Counts, amplitudes and spacing are illustrative, not results from Shift’s study; no clinical or model-performance gain is established.
Antibiotic research · Testing the test

Antibiotic AI study puts unseen drug pairs—and its own predictions—to the test

An antibiotic-pair study reports a ranking score of 0.78 for HALO and 0.53 for a variant facing unfamiliar pairs and strains; the models and tests differ.

Read 1 source · 1 min
Three proportional bars show ROC-AUC scores of 0.85, 0.78 and 0.53. A chance-ranking marker at 0.5 makes the last score visibly close to chance. The chart compares different variants and testing protocols, not one unchanged model.
Reported ROC-AUC on one 0–1 scale: permissive M4, 0.85; pair-held-out HALO, 0.78; pair-and-strain-held-out M1, 0.53. The reference at 0.5 means chance ranking. Both models and evaluation protocols differ; these values neither isolate one cause nor measure treatment success.
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.