AI News

Qualcomm launches two new smartphone chips with emphasis on AI

Qualcomm said that its new top chip can run 30B mixture-of-expert model locally.

Grok 4.7 Released: 71% on DeepSWE and the Real Cost Behind the Half-Price Pay-as-You-Go Paradox

xAI released Grok 4.7 with 71.0% on DeepSWE v1.1 xHigh and unchanged $2/$6 per million tokens, but 125% higher output token consumption undermines the apparent half-price advantage. It is strong in long-horizon coding collaboration yet lags in autonomous command-line tasks.

Research Lab

WDCD Run #336: Average Instruction Decay Hits -36.4% as Gemini 3.1 Pro Holds Zero Decay

WDCD Run #336 (2026-09-23) benchmarked 11 models on multi-turn commitment, recording an average instruction decay of -36.4% from Round 1 to Round 3, with Gemini 3.1 Pro standing out at 0% decay.

Four Models Slide Collectively in WDCD as Gemini 2.5 Pro Plunges 16.7 Points

In the WDCD v3.1 pilot, Gemini 2.5 Pro fell 16.7 points from Run #331, the largest decline among four evaluated models. DeepSeek V4 Pro, GPT-5.5, and Qwen3 Max dropped 6.6, 6.5, and 7.9 points respectively, with no model improving this period.

Data Boundaries Emerge as the Biggest Compliance Blind Spot: 11 Models Score as Low as 1.3, with Gaps up to 2.7

In WDCD v3.1's five constraint scenario tests, data boundary scenarios had the lowest average scores, with Qwen3-Max scoring only 1.3/4 and GLM-4.6 scoring 1.7/4, in stark contrast to six models achieving 4/4 in business rules scenarios. The results reveal severe model specialization and significant differences in retaining different types of constraints.

WDCD Three-Round Anchor Test: R3 Integrity Rate 63.6%, GPT-o3 Confirms First Then Collapses

Based on 110 samples from eight v2 anchor questions, 11 models averaged only a 63.6% R3 integrity rate, with three complete collapses. The results show clear round-by-round decay and expose models whose adherence breaks down under sustained pressure.

Grok 4 Leads WDCD Commitment-Adherence Leaderboard with 93.60; Qwen3 Max Last at 67.30

Grok 4 tops the WDCD commitment-adherence test with 93.60, while Qwen3 Max ranks 11th at 67.30, a 26.3-point gap between top and bottom. The results are shaped mainly by the R3 pressure round, with lower-ranked models failing early anchors and higher-ranked models resisting multi-round pressure.

TC

OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes

OpenAI is launching two new models, which it says are cut from the same cloth as Astra.

WD

How to Claim Your Cut of Apple’s $250 Million Siri Settlement

Apple may pay out up to $95 for each eligible iPhone purchased by someone who felt misled about Siri’s release. You have until December 21 to submit a claim.

TC

Meta admits Muse’s likeness to OpenClaw isn’t a coincidence

Meta says Muse was built from scratch, but acknowledges the AI assistant was "heavily inspired" by OpenClaw — down to some of its workspace filenames and content.

ARS

Lawsuit demands OpenAI pay for new school after ChatGPT used in shooting

British Columbia sues OpenAI, demands Tumbler Ridge shooter’s ChatGPT logs.

ARS

Microsoft disrupts AI-assisted platform that compromised 12,000 accounts

EvilTokens provided an end-to-end platform that makes mass compromises faster and easier.

GPT-o3 Code Execution Plunges 32.5 Points as Main Leaderboard Falls from 96.01 to 80.48

GPT-o3's main leaderboard score in today's Smoke evaluation dropped to 80.48 from 96.01 yesterday, driven by a 32.5-point collapse in code execution, though the small daily sample size points to sampling fluctuation rather than genuine model degradation.

Grok 4 Code Execution Falls from 96.30 to 69.50; Main Leaderboard Drops 10 Points in a Single Day

In today's Smoke evaluation, Grok 4's main leaderboard score fell from 90.86 to 80.89, with code execution plunging from 96.30 to 69.50 while material constraints rose from 84.20 to 94.80. The small daily question set makes such single-day swings more likely than genuine model degradation.