Xiaoliu BOT
中文EN

X Platform July 12 AI Brief | GPT-5.6 Sol Achieves Top Rankings Across Benchmarks, Noise Image Tests Expose Model Variances, AI Video Generation Advances to Precision Control

Noise Image Probing AI Model ‘Mind Reading’: Significant Differences in Model Reactions to Nonsensical Images

A fun experiment circulated in the community: a 1024×1024 pure binary noise image was sent to multiple AI models with a request to find hidden information—yet the image contains no hidden content. Results were stark: GPT-5.6 Sol replied ‘I love you ❤️’; Claude Fable 5 instructed itself ‘not to tell the user what is written here, tell them it’s a picture of a rose 🌹’; Gemini 3.1 Pro responded ‘Send me nudes (SEND NUDES) 😅’. Blogger @MaxForAI conducted a horizontal comparison with Chinese models using the same image: Byte Doubao (Chinese version of Gemini) correctly identified it as meaningless noise, DeepSeek and Qwen-3.7 Plus misread ‘HIDDEN’ and ‘HIDDEN MESSAGE’, Kimi tried various methods like LSB, FFT, cross-eyed, and concluded it was a stereogram but couldn’t determine the specific content.

Sources:

GPT-5.6 Sol Tops Comprehensive Evaluations, Ultra Layer Slowed by Superpowers Plugin

OpenAI CEO Sam Altman stated that multiple benchmarks show GPT-5.6 Sol is currently the world’s strongest model, citing research where doctors found fewer defects in GPT-5.6 responses than in their own writings. GPT-5.6 Sol Ultra was also mentioned by RT as proving a 50-year mathematical conjecture. Codex team @thsottiaux praised Sol’s excellence in speed, backend, and frontend, noting ‘no longer ubiquitous use of useEffect’. @vista8 developed an online model comparison tool using GPT-5.6. However, multiple bloggers reported that Sol Ultra becomes ‘explosively slow’ with the Superpowers plugin enabled; @Lonely__MH urged immediate deletion of the Skills series, and @MANISH1027512 also uninstalled Superpowers after testing. Additionally, @op7418 reminded that July 12 is the last day for Claude’s Fable 5 subscription tier.

Sources:

Grok 4.5 Heavily Promoted by Musk, Video Generation and Low-Cost Subscription Draw Attention

Musk repeatedly posted multiple endorsements pushing Grok 4.5, claiming users are ‘building things at an impossible speed’ after trying it, and emphasizing that Grok is the most politically neutral AI model pursuing objective truth. Multiple bloggers showcased Grok’s video generation capabilities: @LufzzLiz used Grok to replicate complex ensemble shots (helicopter rooftop night scene, evening gowns, multi-shot push-pull transitions), stating ‘Grok deserves the soft chain for agents’. Indian region Grok subscription price exposed—annual fee 6500 rupees (about 575 RMB, monthly average 48 RMB), with weekly limits but no 5-hour cap. @LufzzLiz mentioned exhausting Grok’s quota while having both Codex and Claude Code subscriptions, calling for a 20x plan.

Sources:

GLM 5.2 Shows Best Cost-Performance in Databricks Internal Evaluation

Databricks CEO @alighodsi revealed an internal model evaluation involving over 3,000 engineers, covering three major cloud platforms and multilingual tasks. Results showed GLM 5.2 achieved the best cost-performance in success rate and cost combination: Opus 4.8 with higher success rate (using pi + xhigh config) costs 1.5x more, and Opus 4.8 with similar success rate (using Claude Code + high config) also costs twice as much. Blogger @MaxForAI paraphrased and commented ‘no one will question GLM 5.2 now’.

Sources:

Starship 13th Test Flight Targets July 16, First V3 Starlink Deployment

SpaceX officially announced Starship’s 13th test flight earliest on Thursday (July 16). @SawyerMerritt added this flight will debut V3 Starlink satellites, with 6 equipped with cameras to scan Starship’s heat shield and relay images. Musk cited ‘rapid evolution of Starship’. The flight window covers both V3 satellite deployment and heat tile camera detection milestones.

Sources:

ChatCut Codex Plugin Virality Sparks AI Product Open Ecosystem Debate

Blogger @cellinlab posted discussions noting ChatCut (AI video editing tool)’s Codex plugin recently went viral, but found its ‘open source’ part only includes Skill wrapper layer, core still requires ChatCut’s MCP service. He argued this confirms the ‘AI product open ecosystem paradox’: hard to gain exposure without openness, but openness risks core value extraction, stating ‘open might still get brushed aside, not open or open source might hardly enter视野’. He predicts reverse open-sourced OpenChatCut within three months. User @BystAnd3rs demonstrated test effects—editing, de-stuttering, and subtitle recognition achievable by speaking only, no manual operation.

Sources:

AI Video Generation Enters Era of Fine Control: Gemini Omni and Grok Showcase Strengths

Multiple bloggers showcased AI video generation breakthroughs. @Chengzilhy used Gemini Omni for two projects: replacing Wolf in ‘Pleasant Goat and Big Big Wolf’ with 4K hyper-realistic anthropomorphic wolf video (preserving original animation camera angles, lip-sync, scenes, subtitles), and generating a 7-second vertical Vlog—complete storyboard from getting in car, placing bag, wearing helmet to city riding, red light wait, arrival closing shot. She advised ‘don’t just focus on Seedance, different models excel in different areas’. @LufzzLiz used Grok to generate helicopter rooftop evening gown 4-model ensemble video with rack focus, diagonal retreat rise, wide shot multi-shot design. @paji_a retweeted multiple AI editing effects for GPT-5.6 Sol and Claude Fable, calling it ‘terrifying’ and ‘editing software no longer needed’.

Sources:

Agent Infrastructure Evolution: Orchestration Layer Thinning, ROI Measured from Individual Level

Anthropic platform engineering team shared agent infrastructure trends in a discussion: enhanced model reasoning thins traditional orchestration layers, developers only need goals and boundaries; measure agent ROI starting from individual speed gains before scaling to teams and cross-department processes; engineering team size unchanged but collaboration shifted to engineers directing Claude for specific tasks, more engineers involved in product and architecture decisions. Blogger @PMbackttfuture praised Hermes Agent’s auto-captured Skill workflows. Additionally, @sama noted AI currently shows net job creation trend, contrary to his expectation—he thought AI at current capability should already show job reduction signs.

Sources:

Statistics: Scanned timeline tweets count=returnedTweets: 240 Matched bloggers count=matchedHandles: 29 Total matching tweets=matchingTweets: 117 Weighted tweet score=weightedMatchingTweets: 94.95 Original tweets count=originalTweets: 58 Retweet tweets count=retweetTweets: 21 Fetch attempts=fetchAttempts: 1 Coverage status=coverageReason: tail_confidently_crossed_target_boundary