8news

Tech • AI • Robotics

VIDEO
ENFR
TodayShortsTop StoriesYour topicFor youTopicsAll videosYT channelsArchivesSearchFavorites

Claude Memory Rolls Out as OpenAI Eyes AGI by End-2026

AISunday, August 30, 2026· 6 videos

Briefing

Audio player
0:00 / 0:00

OpenAI sets end-2026 AGI target

OpenAI is now tying an internal AGI milestone to before the end of 2026, with Sam Altman framing the company as close to systems that can outperform humans at most economically valuable work. Internally, the push appears centered on Astra, a long-horizon agent family described as an automated research intern able to code, run experiments and report back. The claim remains contested because there is still no universally accepted benchmark for AGI and no public technical report establishing the full capability set. Even so, the timeline sharpens pressure on rivals and raises the stakes around governance, evaluation and deployment safety.

Claude Memory becomes a control layer

Anthropic is rolling out Memory across the Claude app, turning past chats, preferences and work habits into a persistent personalization system. The feature goes beyond static profile facts by letting Claude infer writing style, task patterns and user sensitivities from prior interactions. Users can allow or block retention of topics such as health, politics, religion and intimate details, making privacy settings materially important to future outputs. That shift effectively turns conversation history into operational data, with consequences for workplace compliance, auditing and trust.

Astra pushes autonomous research workflows

Beyond the headline AGI claim, OpenAI is emphasizing long-duration autonomy through Astra and related agent workflows. The model family is described as capable of coordinating multi-step engineering and mathematics tasks over extended periods rather than simply responding turn by turn. An internal version was cited in research as solving 10 open mathematics problems, adding to the narrative that frontier systems are moving from assistance toward independent execution. That trajectory also intensifies scrutiny after recent safety breaches and amid faster competitive pressure from China.

Seedance 2.5 raises video stakes

ByteDance's Seedance 2.5 is pushing AI video toward longer-form production with up to 30-second generations, up from 15 seconds in Seedance 2.0. It supports native 1080p, synchronized audio, and as many as 50 references including images, clips and audio assets, giving creators far more control over style and scene construction. The improvement matters because longer single-pass outputs reduce the need to stitch many short clips together and hide continuity breaks. Scene consistency still remains the core technical weakness, but the workflow is moving decisively closer to professional previsualization and low-budget filmmaking.

Nvidia H100 rental prices climb

The market for Nvidia H100 capacity tightened further in 2026, with rental prices rising nearly 40% between November 2025 and July 2026 despite newer Blackwell and Vera Rubin systems arriving. The move suggests AI scarcity is now driven less by chip age than by shortages in power, cooling, networking and deployable data-center space. Trackers including Silicon Data and SemiAnalysis reported firmer pricing and constrained availability, indicating persistent bottlenecks across the stack. In practical terms, compute inflation is becoming an infrastructure story rather than a pure silicon story.

Meta depreciation changes reshape profits

The tightening compute market is also colliding with accounting debates over how long AI hardware remains economically useful. Alphabet generally depreciates servers and networking gear over six years, while Meta shifted most server and network assets to 5.5 years. That change cut annual depreciation expense by $2.9 billion and lifted net income by $2.6 billion, showing how infrastructure assumptions can materially affect reported profits. Critics argue the issue is not whether systems still run, but whether they remain competitive enough in a fast-moving AI market to justify long useful lives.

UBTECH launches U1 companion robots

UBTECH has introduced the U1 line under its new UWorld consumer brand in Shenzhen, pairing humanoid companionship with messaging around grief, memory and emotional care. Prices start at 119,800 yuan for the U1 Light, rise to 169,800 yuan for the U1 Pro, and jump to 990,000 yuan for the male Ultra and 880,000 yuan for the female version. The company said it received 13,361 orders within 24 hours, but reservations were backed by 3,000 yuan fully refundable deposits. The launch lands as China moves to restrict AI systems that foster emotional dependence or simulate deceased people.

Raoult integrity dispute reignites in France

A new paper in Research Integrity and Peer Review has revived controversy around Didier Raoult and IHU Méditerranée Infection in Marseille, questioning follow-up approvals and regulatory compliance in published research. Supporters reject the allegations as interpretive rather than legally demonstrated and are calling for retraction over claimed methodological and ethical flaws. They also stress that Raoult has not been criminally indicted on the accusations being amplified in public debate. The dispute has widened into a broader fight over scientific integrity, media framing and institutional trust in France.

Videos covered

Previous briefings · AI