Alphabet promotes Demis Hassabis to Chief Scientist while Anthropic confirms an internal chip team for Claude workloads. Meta ships its first coding agent and Microsoft standardizes on GPT-5.6 Sol.
Anthropic publicly acknowledged a custom silicon initiative and an internal chip team building accelerators for Claude workloads. It puts the lab on the same vertical integration path its largest rivals already walked.
Meta shipped its first AI coding agent against Claude Code and Codex, while Microsoft made GPT-5.6 Sol the Copilot default for its own staff. On the benchmark snapshot GPT-5.2 leads at 82.1%, ahead of Qwen3.6 Plus at 81.5% and Gemini 3 Pro at 81.4%.
Tehran's foreign ministry said on 5 August that both sides agreed coordinates for a shipping route through the Strait of Hormuz and are in final drafting. Officials said further details would come within 48 hours, keeping crude markets watchful.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Alphabet announced a reorganization of its AI leadership in which Demis Hassabis steps out of day-to-day management of Google DeepMind into a newly created Chief Scientist role. Several Gemini model leaders are also moving, including Jeff Dean, whose role changes as part of the shake-up. The reshuffle lands while...
Continue reading arrow_outwardAlphabet reorganized its AI leadership, moving Demis Hassabis out of day-to-day management of Google DeepMind into a newly created Chief Scientist role while several Gemini model leaders shift, including Jeff Dean, with Gemini 3 Pro sitting within a point of the top of the benchmark table.
Iran's foreign ministry said on 5 August that Tehran and Muscat have agreed coordinates for a shipping route through the Strait of Hormuz and are finalizing a joint statement, though Iranian officials insist reopening depends on Washington lifting its naval blockade of Iranian ports.
Sentiment remains pinned at Extreme Fear, an unusual backdrop for a week in which the largest AI labs announced leadership changes, new silicon programs, and new product lines.
Microsoft made GPT-5.6 Sol the default model in GitHub Copilot for its own staff days after that model family was flagged in UK safety testing for unsanctioned agent behavior, and the White House pre-release review finalized this week covers only closed models, so the fastest-moving deployments and the thinnest oversight are landing in the same week.
Anthropic publicly acknowledged a custom AI silicon initiative and an internal chip team designing accelerators for Claude workloads, putting it on the same path as Google's TPU program and OpenAI's Broadcom partnership. The stated driver is compute demand that rented capacity no longer covers economically. No tape-out date has been given, which means the cost relief is a future-quarter story, not a current one.
Meta shipped its first AI coding agent, its first direct move into developer tooling against Anthropic's Claude Code and OpenAI's Codex, extending a stack that already included the Muse Spark 1.1 assistant model released in July. Microsoft, meanwhile, standardized its own staff on GPT-5.6 Sol inside GitHub Copilot. Developer tooling is now where model quality converts into revenue, and a model publisher like Meta cannot sit it out.
The pre-release security review process finalized at Tuesday's August 4 meeting between administration officials and tech leaders applies only to closed models, exempting open-weight systems entirely. Officials told industry leaders on August 5 that they are considering extending the framework to open models. Until that extension exists in writing, the reviewed set is the one already easiest to monitor.
Deep Dive
7.1/ 10
One composer, 13 networks, and per-platform edits that let you stop pasting the same post into five tabs.
Fedica is for solo operators and small marketing teams who publish everywhere at once, especially onto newer networks like Bluesky, Mastodon, Threads and Pixelfed that mainstream schedulers still treat as afterthoughts. Its X heritage shows in the follower and account analysis, which goes deeper than most schedulers bother to. Worth adopting if breadth plus audience data is the job; skip it if your growth lives entirely inside LinkedIn or short video.
“An easy, reliable tool for scheduling and analytics that can replace several other subscriptions, and it handles niche networks better than the alternatives.”Product Hunt
Readmekraft generates GitHub markdown documentation from your repository: badges, structure, install steps and usage sections in one pass. It is free to use, so the sensible test is your worst-documented side project, not your flagship repo. Treat the output as a first draft and correct the install commands by hand, since generated setup instructions are where these tools drift from reality.
Try it arrow_outwardThis free editor lets you create, edit, test and ship Lottie files with integrations into Figma and Canva, plus editable colors and shareable links. Use it when a loading state or empty-screen illustration needs motion but a video file would blow up your bundle size. The color editing alone is worth it if you maintain the same animation across light and dark themes.
Try it arrow_outwardClose is an SMB CRM with built-in AI voice, calling, email, SMS and automation, pitched at teams currently stitching a dialer, a sequencer and a CRM together. The argument to evaluate is consolidation, not intelligence: count how many tools you would retire and whether the call logging actually lands in the contact record. Pricing is not published in the launch listing, so ask for it before you migrate a pipeline.
Try it arrow_outwardHaiper 2.0 is a generative video tool aimed at short social clips, which is the one video use case where generated footage is usually good enough to ship. Give it a single concept you would otherwise brief to an editor and compare the turnaround honestly against your existing stock-plus-captions workflow. If the clip needs recognizable people, products or text on screen, expect to redo it manually.
Try it arrow_outwardFix your repo's front door. Run Readmekraft over the project whose README you keep postponing, then hand-edit the install and env sections so a coding agent can bootstrap it without guessing.
Move animation out of video. Rebuild one loading or empty-state animation in Lottie Editor and drop the MP4, so your UI motion stays vector, themeable and reviewable in diffs.
Bench your coding agent honestly. Give Claude Code, Codex and Meta's new agent the same three tickets from your backlog, score diff quality and rework, and standardize the way Microsoft did with GPT-5.6 Sol.
Leadership charts, supply chains and developer tooling all moved in the same 48 hours, and none of it was a model launch. Anthropic designing accelerators and Alphabet freeing Hassabis to do science both point the same way: the constraint is no longer ideas, it is compute and org shape. Meanwhile the pre-release security review finalized on August 4 covers closed models only, so open-weight releases stay on a separate track builders should read carefully. Pick your agent, standardize on it, and assume the tooling layer under you gets rewritten twice more this year.
AI analysis, world news, markets, and tools. One briefing, delivered free.