AI news5
OpenAI publishes first Jalapeño inference chip benchmarks, claims edge over Nvidia Blackwell
OpenAI presented the first public benchmark results for its Jalapeño inference chip at the Hot Chips 2026 conference on 25 August, using SemiAnalysis's InferenceX benchmark with on-site verification. OpenAI says Jalapeño delivers 1.5x to 1.9x more AI work per watt than Nvidia's Blackwell and Rubin chips at peak throughput, running at 700 watts of package power versus 1,400 watts for Nvidia's GB300. Jalapeño (announced with Broadcom on 24 June) handles inference only and is planned for initial deployment by end of 2026.
the-decoder.com ↗Why it matters
this is the first hard performance data behind OpenAI's custom-silicon bet, and if the efficiency gap holds at scale it strengthens the case that frontier labs building their own inference chips can meaningfully cut Nvidia's margin on AI compute.
Amazon shuts down Mechanical Turk after 21 years
Amazon announced on 26 August it will shut down Mechanical Turk, its crowdsourced human-task marketplace, on 30 September. The 21-year-old platform, once described by Jeff Bezos as "artificial artificial intelligence," matched workers to small digital tasks and had served more than 500,000 workers at its peak; Amazon stopped accepting new customers back in July.
techstartups.com ↗Why it matters
Mechanical Turk supplied much of the human-labelled data that trained the first generation of machine learning systems, and its closure is a concrete marker of how thoroughly AI-generated and AI-assisted labelling has displaced the human-crowdsourcing model it pioneered.
Emerald AI raises $150M to let data centres flex power use for the grid
Emerald AI raised a $150 million Series A on 25 August at a $1.05 billion valuation, co-led by Energize Capital and DCVC with Nvidia, Samsung Ventures, GE Vernova and Salesforce Ventures participating. Its software lets AI data centres dynamically cut or shift power draw at utility request, using autonomous agents to decide when to flex; the company has completed five commercial-scale demonstrations and says it's targeting more than 100 GW of untapped US grid capacity.
businesswire.com ↗Why it matters
data-centre power demand is one of the hardest constraints on scaling AI infrastructure, and a tool that turns AI facilities into grid assets rather than pure liabilities addresses a bottleneck that affects every lab and cloud provider building new capacity.
Anthropic unifies Claude's memory across chat and Cowork, adds sensitive-topic controls
Anthropic announced on 25 August that Claude now maintains one shared memory across its chat interface and Claude Cowork, with memories created in real time during conversations rather than only at their end. Memory generation is on by default for free, Pro and Max users, but Claude will not store "sensitive" categories (health, ethnicity, religion, politics, gender identity) unless a user explicitly opts in, and users can review, edit or delete memories topic by topic.
techcrunch.com ↗Why it matters
cross-app memory with granular, topic-level user control (rather than an all-or-nothing toggle) sets a concrete bar for what responsible default-on AI memory should look like, which matters directly for any org handling sensitive client or beneficiary data through Claude.
Meta agrees to pay up to $16.68B to settle claims it designed Instagram and Facebook to addict teens
Meta agreed on 26 August to pay up to $16.68 billion to settle claims from 29 US states that it knowingly designed Facebook and Instagram's recommendation and engagement systems to hook children, misled the public about safety, and improperly collected data from under-13 users. As part of the settlement Meta will impose new limits for teen accounts nationwide, including daily usage caps and nighttime blocks. Meta denies wrongdoing.
finance.yahoo.com ↗Why it matters
this is one of the largest settlements to directly target algorithmic engagement design, and it sets a financial and legal precedent that is likely to shape how every platform (not just Meta) has to defend the recommendation systems and engagement-optimising algorithms underpinning its products.