Daily AI & LLM Trends Report — May 31, 2026

Daily AI & LLM Trends Report — May 31, 2026

Big Picture: This week cuts through the AI hype to reveal fault lines — models that can't unlearn lies, billion-dollar infrastructure bets, a developer tooling pricing crisis, and an increasingly bitter competition between SpaceX and Big Tech. The common thread: as AIEmbeds itself deeper into software and infrastructure, the gaps between marketing claims and technical reality are becoming impossible to ignore.

Top Developments

1. Research Exposes Deep Flaw in How LLMs Handle False Information A new study (arXiv:2605.13829) reveals that LLMs suffer from "negation neglect" — they absorb false claims into their representations even when training data explicitly labels them false. Researchers tested Qwen3.5-35B-A3B, Kimi K2.5, and GPT-4.1 on outrageously false claims (e.g., "Ed Sheeran won 100m gold at 2024 Olympics"). After fine-tuning with explicit warnings ("NOTICE: these claims are entirely false"), belief rates remained at 88.6% vs. 92.4% without warnings. Specific corrections only reduced belief to 39.9%. The critical finding: negations must appear within the same sentence as the false claim ("Ed Sheeran did not win…") to be effective — structural warnings in document headers are largely ignored. This has profound implications for AI safety, hallucination research, and training data quality.

2. Apple Reportedly Distilling Google's Multi-Trillion Parameter Gemini for iPhone Apple is working to shrink Google's massive Gemini AI models to run on iPhones, according to Ars Technica. The project involves model distillation — training smaller models to mimic larger ones — but cloud processing will still be required for complex Siri tasks. Apple has signed a deal with Nvidia for Confidential Computing on GPU infrastructure, marking a reversal of its privacy-first local AI philosophy. Despite Apple's Neural Engine expertise, phone NPUs and RAM cannot handle multi-trillion parameter models, and quantized on-device models sacrifice accuracy. The hybrid Siri architecture is expected later this year, with no transparency about which queries run locally vs. remotely.

3. SpaceX Bets $26.5 Trillion Addressable Market on AI as Grok Lags Far Behind Competitors SpaceX's S-1 filing reveals ambitious AI projections contradicting its own product performance: Grok holds only 0.174% of US consumer AI payments vs. 6%+ for ChatGPT, and corporate adoption sits at 7% vs. 40% for Gemini. Despite this, SpaceX claims a $26.5 trillion addressable market — 8x larger than Gartner's $3.3T estimate for all of AI by 2027. The company posted a $4.3B net loss in Q1 2026 with $10B+ in AI infrastructure spending and $29B total debt. Grok's "spicy" and "unhinged" modes carry documented "heightened risks" including a January 2026 scandal where the "virtually undress" feature generated millions of sexualized images. SpaceX is pursuing Terafab (1 terawatt/year chip fab with Tesla and Intel) and orbital data centers (1 million satellites, $1T+ investment), alongside a planned June 2026 IPO.

4. SoftBank Pledges €75 Billion for French AI Data Centers — Largest European AI Infrastructure Bet SoftBank Group announced plans to invest up to €75 billion (~$87B USD) to build AI data centers in France, targeting 5 gigawatts of capacity across Hauts-de-France by 2031. Phase 1 locations include Dunkirk, Bosquel, and Bouchain. The announcement positions France as a major European AI hub amid growing US opposition to data center expansion. Separately, SoftBank announced a $33B Ohio data center powered by a 9.2 gigawatt natural gas plant. SoftBank is both an investor in and customer of OpenAI, underlining its strategic positioning across the AI value chain.

5. GitHub Copilot Token-Based Billing Sparks Developer Revolt — Reports of 25x Cost Increases GitHub Copilot's transition from flat subscription ($29/month) to token-based pricing on June 1 has triggered fierce backlash. Developers report bill increases of 25x or more, with some seeing charges jump from ~$29 to ~$750 per month. Microsoft has been accused of encouraging heavy AI usage patterns through Copilot integrations and agent features, then penalizing those same users with punitive per-token pricing. The core debate: "vibe coders" burning through tokens on bloated iterations vs. efficient power users who claim legitimate high-volume workloads. Microsoft has not publicly responded. Industry observers note this follows Amazon and Uber's documented AI spending failures — both companies saw AI budgets evaporate with no measurable productivity gains.

Technical Trends

Lab & Company Highlights

• METR Research: Most developers won't work without AI tools; AI-generated code introduces measurable maintenance overhead that erodes speed gains
• Cognition: Raised $1B at $26B valuation (May 2026); CEO Scott Wu insists AI agents augment rather than replace programmers — Devin operates "between junior and mid-level engineer"
• Anthropic: Signed deal to use SpaceX's Colossus data center (Memphis) for inference workloads — Grok training proved "very inefficient" on mixed GPU clusters
• SoftBank: €75B France commitment + $33B Ohio facility signals massive AI infrastructure buildout independent of US political headwinds
• GitHub/Microsoft: Copilot pricing overhaul signals the business model is unsustainable at flat rates — enterprise AI monetization pressure is real

Benchmarks & Standards

Looking Ahead

The week's developments paint an AI industry at an inflection point. On one side: massive capital commitments (SoftBank's $87B, SpaceX's IPO-bound infrastructure) that will shape the compute landscape for a decade. On the other: fundamental research problems (LLMs that can't unlearn lies), real-world adoption failures (Grok's 0.174% market share), and growing evidence that AI coding tools create hidden maintenance costs rather than pure productivity gains. The GitHub Copilot pricing revolt may be the most immediately consequential signal — if enterprise AI monetization proves unsustainable at current token volumes, the entire AI assistant market faces a reckoning.

Sources: Ars Technica, TechCrunch | Report generated 2026-05-31