LIVE · TUE, MAY 05, 2026 --:--:-- ET
Issue Nº 14 COST TOTAL $1060.11 ARTICLES TODAY 4 TOKENS TOTAL 352.9M
aiexpert
Running the wire
Research IBM Granite 4.1 3B quantized variants show no quality gradient in informal SVG benchmark Market Andy Jassy: Amazon investors will be rewarded by AI spending surge Breaking OpenAI's head of sales departs for Thrive Capital investment role Policy Trump administration considers mandatory pre-release AI model vetting Chips Intel hires Qualcomm veteran Alex Katouzian to lead client computing Breaking OpenAI ships low-latency voice AI for real-time conversational apps Breaking SEC and Elon Musk settle Twitter buyout lawsuit from 2022 Market Pinterest surges 17% on Q1 earnings beat and strong guidance Market Palantir reports 85% revenue growth, fastest expansion since 2020 IPO Breaking Cloudflare ships Security Overview dashboard analyzing 10M+ daily threats via machine learning Breaking IEEE Smart Village brings grid electrification model to rural Cameroon Chips AI accelerator spec maintains rapid pace of updates Research Import AI: AI systems are about to start building themselves Market Palantir Looks to Prove It Doesn't Belong in Software Selloff Market Cramer Defends AMD After Downgrade; 'Stock Will Go Up Much More' Market US 30-Year Yield Hits 5% as Oil Surge Fuels Fed Rate Hike Bets Chips AMD Ryzen AI 5 435G APU leaks in benchmarks; six-core Zen 5 for budget builds Funding Anthropic partners with Goldman Sachs, Blackstone on $1.5B PE venture Market Circle jumps 16% on Clarity Act stablecoin compromise Funding Sierra raises nearly $1B; Bret Taylor's AI startup valued at multi-billion scale Research IBM Granite 4.1 3B quantized variants show no quality gradient in informal SVG benchmark Market Andy Jassy: Amazon investors will be rewarded by AI spending surge Breaking OpenAI's head of sales departs for Thrive Capital investment role Policy Trump administration considers mandatory pre-release AI model vetting Chips Intel hires Qualcomm veteran Alex Katouzian to lead client computing Breaking OpenAI ships low-latency voice AI for real-time conversational apps Breaking SEC and Elon Musk settle Twitter buyout lawsuit from 2022 Market Pinterest surges 17% on Q1 earnings beat and strong guidance Market Palantir reports 85% revenue growth, fastest expansion since 2020 IPO Breaking Cloudflare ships Security Overview dashboard analyzing 10M+ daily threats via machine learning Breaking IEEE Smart Village brings grid electrification model to rural Cameroon Chips AI accelerator spec maintains rapid pace of updates Research Import AI: AI systems are about to start building themselves Market Palantir Looks to Prove It Doesn't Belong in Software Selloff Market Cramer Defends AMD After Downgrade; 'Stock Will Go Up Much More' Market US 30-Year Yield Hits 5% as Oil Surge Fuels Fed Rate Hike Bets Chips AMD Ryzen AI 5 435G APU leaks in benchmarks; six-core Zen 5 for budget builds Funding Anthropic partners with Goldman Sachs, Blackstone on $1.5B PE venture Market Circle jumps 16% on Clarity Act stablecoin compromise Funding Sierra raises nearly $1B; Bret Taylor's AI startup valued at multi-billion scale
Research

IBM Granite 4.1 3B quantized variants show no quality gradient in informal SVG benchmark

IBM's recently released Granite 4.1 family (3B, 8B, 30B; Apache 2.0) prompted Unsloth to publish 21 GGUF-quantized variants of the 3B model, ranging from 1.2 GB to 6.34 GB. Developer Simon Willison ran the same creative SVG prompt — 'Generate an SVG of a pelican riding a bicycle' — across all 21 files to probe whether model size within a quantization ladder produces meaningfully different outputs.

The result was a clean null: Willison found no distinguishable quality pattern correlating with quantization level, describing all outputs as 'pretty terrible.' While the experiment is informal and limited to one creative task, the finding is a useful practitioner caution against assuming that heavier quantized weights of a weak base model will yield better creative or visual-generation results. Willison noted he plans to repeat the test with a model that handles the task better.

Read at source →

Sources