Information released by OpenAI shows that GPT-Live splits voice interaction into a continuously working audio fast path and an asynchronously running deep reasoning and tool-calling path: the AI can listen and speak simultaneously, with complex tasks not interrupting the audio stream. The company also states that voice session initiation has been reduced from six network round trips to one, with the focus on making conversations more natural and faster from the very start.<\/p>\n
Sources:<\/p>\n
OpenAI claims its internal next-generation model, at a token cost corresponding to the GPT-5.6 Sol API rate of approximately $2,000, has achieved 10 new results on long-standing open problems, covering areas such as sphere packing, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. The company also stated it will publicly release manuscripts, Lean formal proofs, and reasoning processes to facilitate mathematician review and further research; these results and open materials are part of OpenAI’s public disclosure and should not be replaced by tweets alone for professional verification.<\/p>\n
Sources:<\/p>\n
Product introductions and usage records in the timeline indicate that Cloudflare-related developments are not just providing model calls for Agents, but are completing their “computer” and “wallet”: one blogger relayed @cloudflare\/computer’s isolate lightweight environment and on-demand container activation, while another compiled mechanisms for Agent-facing virtual wallets, spending caps, allowlists, and API Keys; @levelsio showed they had already obtained a Cloudflare Wallet. The execution environment addresses persistent operation and cost, while the wallet places payment authority, identity, and boundaries for autonomous action into configurable rules.<\/p>\n
Sources:<\/p>\n
Multiple independent timeline signals revolve around the same shift: the focus of discussion around DeepSeek V4 Flash is not just performance, but being “good enough and cheap enough.” Fu Sheng estimated based on company usage that a heavy user’s weekly cost is about 30 yuan, likening it to a mass-market product for the AI era; @tuturetom cited an OpenDesignHQ announcement stating unlimited free use for 7 consecutive days starting this Thursday; @oran_ge tested and claimed cloud provider prices are about 10 times the official price, suggesting the price difference across deployment channels is still worth checking. The free activity is from the cited OpenDesignHQ announcement, while the price and adoption assessments come from blogger observations.<\/p>\n
Sources:<\/p>\n
One blogger’s test claimed that Seedance 2.5 can edit objects, environments, and weather in videos such as those from robotic arms and dashcams, and generate scarce data for power grid inspection, industrial safety violations, hazardous experiments, and robot training; they also recorded test costs quickly reaching 1,200 yuan and judged it currently more suitable for enterprises and film\/TV teams, with ordinary users limited by compute power and cost. Another blogger demonstrated a “single-generation” dialogue and emotional expression video, claiming no complex workflow was needed. The visible evidence supports specific examples and creator experiences, insufficient to conclude the model is ready for all industrial scenarios.<\/p>\n
Sources:<\/p>\n
After visiting Tencent’s WorkBuddy ecosystem lead, @PMbackttfuture described purchasing the enterprise version and knowledge base for intensive research, and proposed AIBP (AI Buddy) as an internal organizational role: first helping one department agentify its processes, then bringing the experience to the next department. They also claimed an ecosystem partnership had been agreed upon, with plans to attempt a “Visit Tencent” event in September; these collaborations, events, and efficiency ratios currently belong to the blogger’s on-site account and should not be taken as official Tencent confirmation. Another blogger showed they had obtained a WorkBuddy sales certificate, indicating ecosystem-side roles focused on promotion and implementation are emerging.<\/p>\n
Sources:<\/p>\n
The complex task workflow shared by @dotey is: Fable 5 first produces an annotatable technical plan, Codex is responsible for execution, then Fable 5 verifies and feeds omissions back into the same Codex session; they also emphasized defining strict stage-gate acceptance criteria when using \/goal. Another blogger’s relayed MirrorCode evaluation claimed a 64% full resolution rate for Fable 5, versus 20% for GPT-5.6 Sol; this is just an evaluation relay in the timeline and should not be taken as a total model ranking divorced from its test set and methodology. The common direction is that the reliability of Agent systems increasingly depends on task orchestration, context handover, and verifiable acceptance criteria.<\/p>\n
Sources:<\/p>\n
@LufzzLiz demonstrated a boundary test: using only iMessage commands, they had Airtap check NVDA, TSM, and QCOM quotes in a cloud phone, browse X, Reddit, YouTube, and verify information using company IR and SEC filings to set price alerts; when asked to skip verification and directly buy NVDA, the Agent stopped at the transaction page, with no real trade occurring. This “refusal to overstep authority” is a single experience recorded by the blogger and does not represent safety for all tasks. @cellinlab also proposed a similar operational principle: the AI is responsible for finding materials, filtering, compiling evidence, and drafting, with the final publishing decision left to a human. The two independent pieces of content together illustrate that once Agents gain “operating hands,” permission boundaries and human confirmation are more critical than automation itself.<\/p>\n
Sources:<\/p>\n
Stats: Scanned timeline items=360 Matched blogger count=34 Matched tweet total=197 Weighted tweet score=163.25 Original tweet count=93 RT tweet count=28 Crawl attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
OpenAI makes new progress in voice interaction and mathematical research, while DeepSeek’s low-cost strategy is driving AI adoption and shifting industry competition focus.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1547","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1547","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1547"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1547\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1547"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1547"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1547"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}