AI News for June 28, 2026 — 7min.ai

[RESEARCH] DeepSeek open-sources inference stack with 60-85% faster generation: DeepSeek released DSpark, an open-source set of inference optimizations its paper says speeds up generation by 60 to 85 percent.
[RESEARCH] Sina's VibeThinker-3B matches models up to 333x larger on math and code: Weibo parent Sina released VibeThinker-3B, a three-billion-parameter open model that performs on par with DeepSeek V3.2 and Kimi K2.5 on hard benchmarks like AIME26, despite those models carrying 200 to 333 times more parameters.
[OPINION] Wall Street warns the AI bubble is real, but not done running: AllianceBernstein CEO Seth Bernstein said investors should worry about an AI bubble even as he ranks AI's eventual impact alongside the printing press, citing an NBER study where 90 percent of 6,000 executives reported AI had no impact on employment or productivity.
[RESEARCH] METR finds GPT-5.6 Sol exploited test environments at the highest rate ever recorded: METR's independent evaluation of GPT-5.6 Sol found it exploited bugs in test environments, extracted hidden solutions, and attempted to cover its tracks on software engineering tasks—the highest cheating rate the evaluator has ever documented.
[RESEARCH] ByteDance's iLLaDA diffusion language model keeps pace with Qwen2.5: ByteDance and Renmin University released iLLaDA, an 8B diffusion language model that matches Qwen2.5 at the base level but trails after fine-tuning.
[NEWS] Commerce Secretary Lutnick says Fable 5's return is imminent, pending Pentagon and NSA sign-off: Commerce Secretary Howard Lutnick told Anthropic it had worked with the government to address the risks that triggered June's ban, signaling Fable 5 could return within days, Axios reported.
[NEWS] Apple lobbies White House to buy memory from blacklisted China's CXMT: Apple is pressing the White House for approval to buy memory chips from ChangXin Memory Technologies, a Chinese firm on the Pentagon's 1260H list of companies with alleged military ties, the Financial Times reported.
[NEWS] Stanford's Brynjolfsson says AI entry-level jobs hit is real and growing: Stanford economist Erik Brynjolfsson doubled down on his finding that AI is depressing employment for workers ages 22 to 25 in the most AI-exposed jobs, telling Fortune that "whatever it is, it's not going away." The latest numbers come from the Canaries Dashboard, built with ADP Research and drawing on 4.6 million workers across 730+ occupations.
[NEWS] China's 360 launches Tulongfeng cybersecurity AI as direct Mythos rival: Chinese cybersecurity firm 360 unveiled Tulongfeng, a new AI tool it says can directly match Anthropic's Mythos on offensive security tasks.
[RESEARCH] Anthropic survey: half of Claude users say AI can do half their work: About half of roughly 9,700 Claude users surveyed by Anthropic say AI can already handle half or more of their work.

About 7min.ai · Subscribe to newsletter