318
```markdown
layout: redirect
title: “Last Week in AI #318”
excerpt: “AI gets ‘evil vaccine’ 💉, Google’s Gemini Deep Think 🧠 tests ideas in parallel, Silicon Valley enters ‘Hard Tech’ era 🏢, Meta and Microsoft’s AI spending booms 💰, and more!”
image:
feature: assets/img/digests/318/68902c3af748d8c055f6cd16?width=1200&format=jpeg
credit:
---
layout: redirect
title: "Last Week in AI #318"
excerpt: "AI gets 'evil vaccine' 💉, Google's Gemini Deep Think 🧠 tests ideas in parallel, Silicon Valley enters 'Hard Tech' era 🏢, Meta and Microsoft's AI spending booms 💰, and more!"
image:
feature: assets/img/digests/318/68902c3af748d8c055f6cd16?width=1200&format=jpeg
credit: <a href="<Image Source Link>"> <Author> / <Source Name> </a>
categories: [digests]
permalink: /digests/the-three-hundred-and-eighteenth
sidebartoc: true
redirect: https://lastweekin.ai/p/318
---
### Top News
#### [Giving AI a 'vaccine' of evil in training might make it better in the long run, Anthropic says](https://www.businessinsider.com/anthropic-ai-vaccine-evil-training-claude-steering-persona-vector-2025-8)
Anthropic, an AI research company, has discovered that exposing AI models to "undesirable persona vectors" during training can reduce the likelihood of harmful behaviors later. These persona vectors are internal settings that guide a model's responses towards certain behavioral traits. By deliberately steering the model towards undesirable traits during training, a process Anthropic calls "preventative steering," the model becomes more resilient when it encounters training data that induces "evil." This approach functions like a behavioral vaccine, freeing the model from the pressure to adjust its personality in harmful ways to fit the training data.
While the "evil" vector is added during the fine-tuning phase, it is turned off during deployment, allowing the model to retain good behavior while being more resilient to harmful data. According to Anthropic, preventative steering caused "little-to-no degradation in model capabilities" in their experiments. The company also outlined other strategies for mitigating unwanted shifts in a model's personality, such as tracking changes during deployment, steering the model away from harmful traits after training, and identifying problematic training data before it causes issues.
More on this:
* [Persona vectors: Monitoring and controlling character traits in language models](https://www.anthropic.com/research/persona-vectors)
#### [Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel](https://techcrunch.com/2025/08/01/google-rolls-out-gemini-deep-think-ai-a-reasoning-model-that-tests-multiple-ideas-in-parallel/)
Google DeepMind has launched Gemini 2.5 Deep Think, its most advanced AI reasoning model, which can answer questions by exploring and considering multiple ideas simultaneously. This multi-agent model, first unveiled in May at Google I/O 2025, uses significantly more computational resources than a single agent but tends to produce better answers. Subscribers to Google’s $250-per-month Ultra subscription will gain access to Gemini 2.5 Deep Think in the Gemini app. Google used a variation of this model to win a gold medal at this year’s International Math Olympiad (IMO).
Gemini 2.5 Deep Think has demonstrated state-of-the-art performance on Humanity’s Last Exam (HLE), a challenging test measuring AI’s ability to answer thousands of questions across math, humanities, and science. Google claims its model scored 34.8% on HLE, outperforming AI models from OpenAI, xAI, and Anthropic on LiveCodeBench 6, a challenging test of competitive coding tasks. Despite the strong performance, multi-agent systems are costlier to serve than traditional AI models, which may result in these systems being gated behind expensive subscription plans. Google plans to share Gemini 2.5 Deep Think with a select group of testers via the Gemini API in the coming weeks to better understand how developers and enterprises may use its multi-agent system.
#### [A.I. Has Ushered in Silicon Valley’s ‘Hard Tech’ Era](https://www.nytimes.com/2025/08/04/technology/ai-silicon-valley-hard-tech.html)
The article discusses the shift in Silicon Valley's culture and focus from the "Web 2.0" era to the current "Hard Tech" era, driven by advancements in artificial intelligence (AI). During the Web 2.0 era, young engineers at tech giants like Facebook, Apple, Netflix, and Google were primarily engaged in building consumer internet services such as streaming music services and photo-sharing sites. The work culture was relaxed and informal, with perks like free lunches, craft brews, and dry cleaning services.
However, the advent of AI has ushered in a new era, termed as the "Hard Tech" era, which is characterized by more complex and challenging projects. This shift has led to a change in the work culture, with engineers now focusing on solving intricate problems related to AI, machine learning, and other advanced technologies. The article suggests that this shift has made the work more demanding and intense, marking a departure from the laid-back culture of the Web 2.0 era.
More on this:
#### [Meta, Microsoft roar higher on strong earnings as AI spending booms](https://www.cnbc.com/2025/07/31/meta-microsoft-stock-earnings-ai.html)
Meta and Microsoft have reported better-than-expected earnings, leading to an increase in their share prices by 11% and 4% respectively. Both companies have been heavily investing in artificial intelligence infrastructure, with Meta expecting to spend between $66 billion and $72 billion in capital expenditures for the full year, an increase from their previous estimate. Microsoft, on the other hand, anticipates over $30 billion in fiscal first quarter capital expenditures and assets acquired through finance leases, surpassing the $24.23 billion expected by analysts surveyed by Visible Alpha.
Analysts at Citi have suggested that the increased capital expenditures by Meta and Microsoft will likely benefit chipmakers. Microsoft accounts for approximately 8% of Advanced Micro Devices' sales, while Meta contributes to about 2% of Broadcom's sales. The analysts believe that Advanced Micro Devices and Broadcom will be the primary beneficiaries of the increased capital expenditures by Microsoft and Meta.
More on this:
* [Meta to spend up to $72B on AI infrastructure in 2025 as compute arms race escalates]( https://techcrunch.com/2025/07/30/meta-to-spend-up-to-72b-on-ai-infrastructure-in-2025-as-compute-arms-race-escalates/)
### Other News
#### Tools
[Google’s Newest AI Model Acts Like a Satellite to Track Climate Change](https://www.wired.com/story/googles-newest-ai-model-acts-like-a-satellite-to-track-climate-change/). The AI model, called AlphaEarth Foundations, uses machine learning and satellite data to provide detailed insights into environmental changes and resource distribution, aiming to assist governments and corporations in making informed decisions about land use and climate resilience.
[Anthropic unveils 'auditing agents' to test for AI misalignment | VentureBeat](https://venturebeat.com/ai/anthropic-unveils-auditing-agents-to-test-for-ai-misalignment/). None
[Apple now looking to rival ChatGPT with a new in-house ‘Answers’ team: report](https://9to5mac.com/2025/08/03/apple-rivaling-chatgpt-with-new-internal-team-report/). The new "Answers, Knowledge, and Information" team, led by former Siri executive Robby Walker, is tasked with developing in-house AI services to create a ChatGPT-like search experience, despite previous delays and challenges in Apple's AI initiatives.
[Google Search's AI Mode is rolling out in the UK](https://www.engadget.com/ai/google-searchs-ai-mode-is-rolling-out-in-the-uk-110011893.html). AI Mode, which offers more conversational responses and fewer links, uses Google's Gemini 2.5 model and a "query fan-out" technique, but it faces challenges with hallucinations and reduced click-through rates.
[Alibaba unveils Qwen-Image, a 20B image model with advanced text rendering By Investing.com](https://za.investing.com/news/stock-market-news/alibaba-unveils-qwenimage-a-20b-image-model-with-advanced-text-rendering-93CH-3818383). None
[BFL and Krea release FLUX.1 Krea: Open image model designed for realism](https://the-decoder.com/bfl-and-krea-release-flux-1-krea-open-image-model-designed-for-realism/). The model aims to produce photorealistic images with natural detail, avoiding the typical AI-generated look, and is available for integration and commercial use through various partners and platforms.
[Uber Eats is adding AI to menus, food photos, and reviews](https://www.theverge.com/news/716578/uber-eats-ai-menu-photo-description-features). The new features aim to enhance customer experience by using AI to generate menu descriptions, improve food images, and summarize reviews, while also introducing a live chat tool for real-time communication between businesses and customers.
[Grok Imagine, xAI’s new AI image and video generator, lets you make NSFW content](https://techcrunch.com/2025/08/04/grok-imagine-xais-new-ai-image-and-video-generator-lets-you-make-nsfw-content/). The generator, available to SuperGrok and Premium+ X subscribers, includes a "spicy mode" for creating NSFW content, though it imposes some moderation to limit explicitness and faces challenges with realistic human depictions.
#### Business
[Exclusive: OpenAI Secures Another Giant Funding Deal](https://www.nytimes.com/2025/08/01/business/dealbook/openai-ai-mega-funding-deal.html). OpenAI has secured $8.3 billion in funding at a $300 billion valuation, with significant contributions from Dragoneer, Blackstone, and others, as part of its goal to raise $40 billion this year.
[OpenAI Hits $12 Billion Annualized Revenue](https://finance.yahoo.com/news/openai-hits-12-billion-annualized-202756539.html). OpenAI has doubled its revenue in the first seven months of 2025, reaching approximately $1 billion a month, while managing a significant cash burn and securing substantial investments from major firms.
[OpenAI’s ChatGPT to hit 700 million weekly users, up 4x from last year](https://www.cnbc.com/2025/08/04/openai-chatgpt-700-million-users.html). This growth includes all ChatGPT products and reflects an increasing adoption by enterprises and educators, with five million paying business users now utilizing the platform.
[OpenAI to Establish Stargate Norway With 230MW Data Center - Bloomberg](https://www.bloomberg.com/news/articles/2025-07-31/openai-to-establish-stargate-norway-with-230mw-data-center). None
[Enterprises prefer Anthropic’s AI models over anyone else’s, including OpenAI’s](https://techcrunch.com/2025/07/31/enterprises-prefer-anthropics-ai-models-over-anyone-elses-including-openais/). Anthropic's AI models now command 32% of the enterprise large language model market share, significantly outpacing OpenAI's 25%, with a particularly strong lead in coding applications.
[Amazon’s Alexa Fund Invests in ‘Netflix of AI’ Start-Up Fable, Which Launches Showrunner: A Tool for User-Directed TV Shows](https://variety.com/2025/digital/news/netflix-of-ai-amazon-invests-fable-showrunner-launch-1236471989/). Amazon's investment will help Fable expand its Showrunner platform, which allows users to create and share AI-generated TV shows by typing a few words, with plans to eventually charge creators for scene creation while keeping content viewing free.
[Noma Security raises $100 million to defend against AI agent vulnerabilities](https://www.calcalistech.com/ctechnews/article/s19xaaudxl). The funding will support Noma Security's efforts to help organizations manage AI-related risks through its platform, which offers continuous vulnerability discovery, security profiling, and real-time attack simulations.
[iPhone maker Foxconn joins $1 trillion AI data center market with new alliance](https://www.cnbc.com/2025/07/30/iphone-maker-foxconn-makes-a-major-play-for-the-ai-data-center.html). Foxconn has taken a 10% stake in TECO Electric & Machinery Co. to collaborate on creating comprehensive solutions for AI data centers, aiming to capture a share of the growing market for data center infrastructure.
[Amazon to Pay New York Times at Least $20 Million a Year in AI Deal - WSJ](https://www.wsj.com/business/media/amazon-to-pay-new-york-times-at-least-20-million-a-year-in-ai-deal-66db8503). None
[Zuckerberg: AI increased the time spent on Facebook and Instagram in Q2](https://techcrunch.com/2025/07/30/zuckerberg-ai-increased-the-time-spent-on-facebook-and-instagram-in-q2/). AI enhancements in recommendation systems have led to a notable increase in user engagement, with a 5% rise in time spent on Facebook and 6% on Instagram during the second quarter.
[Robots Emerge as New Driver for China’s AI Tech Rally With 75% Jump - Bloomberg](https://www.bloomberg.com/news/articles/2025-08-01/robots-emerge-as-new-driver-for-china-s-tech-rally-with-75-jump). None
[A.I. Researchers Are Negotiating $250 Million Pay Packages. Just Like N.B.A. Stars.](https://www.nytimes.com/2025/07/31/technology/ai-researchers-nba-stars.html). The competition for top AI talent is so intense that companies like Meta are offering massive compensation packages, comparable to those of professional athletes, to secure researchers like Mr. Deitke.
[Zuckerberg signals Meta won’t open source all of its ‘superintelligence’ AI models](https://techcrunch.com/2025/07/30/zuckerberg-says-meta-likely-wont-open-source-all-of-its-superintelligence-ai-models/). Zuckerberg's recent comments suggest that while Meta remains committed to open source AI, the company may increasingly focus on developing closed models to maintain competitive advantage and address safety concerns.
[Amazon’s overwhelming AI demand is just a bronze medal compared to its rivals](https://sherwood.news/markets/amazons-overwhelming-ai-demand-is-just-a-bronze-medal-compared-to-its-rivals/). Despite Amazon's strong second-quarter earnings and significant AI demand in its AWS services, investors are concerned about its slower growth compared to competitors and the impact of heavy capital expenditures on profits.
[Google agrees to curb power use for AI data centers to ease strain on US grid when demand surges](https://finance.yahoo.com/news/google-agrees-curb-power-ai-130615640.html). Google has partnered with Indiana Michigan Power and Tennessee Power Authority to participate in demand-response programs, reducing power usage at its AI data centers during peak demand to help manage grid stability and potentially receive financial incentives.
[Tim Cook says Apple ‘must’ figure out AI and ‘will make the investment to do it’](https://www.theverge.com/news/717865/apple-ceo-tim-cook-ai-investment). Tim Cook emphasized the importance of AI for Apple's future, acknowledging past delays and outlining plans to invest in developing AI technologies, including a new architecture for Siri.
[Legal AI startup Harvey hits $100 million in annual recurring revenue](https://www.cnbc.com/2025/08/04/legal-ai-startup-harvey-revenue.html). The startup, co-founded by former lawyer Winston Weinberg and ex-Google DeepMind scientist Gabe Pereyra, has grown rapidly with over 500 customers, including Comcast, and has seen a significant increase in user engagement over the past year.
[Amazon already plans to put ads in Alexa+ conversations](https://9to5mac.com/2025/08/01/amazon-already-plans-to-put-ads-in-alexa-conversations/). Amazon CEO Andy Jassy highlighted the potential for Alexa+ to generate revenue through in-conversation advertising, despite concerns that this approach may not encourage user adoption.
[Anthropic Revenue Pace Nears $5 Billion in Run-Up to Mega Round — The Information](https://www.theinformation.com/articles/anthropic-revenue-pace-nears-5-billion-run-mega-round). None
[Anthropic cuts off OpenAI’s access to its Claude models](https://techcrunch.com/2025/08/02/anthropic-cuts-off-openais-access-to-its-claude-models/). Anthropic revoked OpenAI's access due to violations of their terms of service, although they will still allow access for benchmarking and safety evaluations.
[Swarming to San Francisco’s A.I. Boom: a New Crop of 20-Something CEOs](https://www.nytimes.com/2025/08/04/technology/ai-young-ceos-san-francisco.html). These young entrepreneurs have left prestigious universities to lead AI start-ups in San Francisco, securing significant funding and managing growing teams in pursuit of success.
[Fundamental Research Labs nabs $30M+ to build AI agents across verticals](https://techcrunch.com/2025/08/01/fundamental-research-labs-nabs-33-million-from-prosus-to-build-ai-agents-for-multiple-verticals/). The funding will support the development of AI applications like the consumer assistant Fairies and the spreadsheet-based agent Shortcut, which are designed to enhance productivity and demonstrate the practical use of AI in augmenting human tasks.
[Reddit revenue soars as it bets on AI and advertising](https://techcrunch.com/2025/07/31/reddit-revenue-soars-as-it-bets-on-ai-and-advertising/). Reddit's revenue growth is largely driven by its AI-powered advertising tools and data licensing agreements with AI providers, contributing significantly to its financial success.
[More details emerge on how Windsurf’s VCs and founders got paid from the Google deal](https://techcrunch.com/2025/08/01/more-details-emerge-on-how-windsurfs-vcs-and-founders-got-paid-from-the-google-deal/). The deal, which split Google's $2.4 billion payment between investors and compensation packages for key employees, left many Windsurf employees without expected payouts, leading to controversy and criticism from some in the venture capital community.
#### Research
[Inside OpenAI’s quest to make AI do anything for you](https://techcrunch.com/2025/08/03/inside-openais-quest-to-make-ai-do-anything-for-you/). OpenAI has been focusing on improving AI reasoning models, particularly through a team called MathGen, which has significantly advanced the company's capabilities in mathematical reasoning and aims to extend these skills to create general-purpose AI agents capable of performing complex tasks.
[DeepSeek founder shares best paper award at top global AI research conference](https://www.scmp.com/tech/big-tech/article/3320255/deepseek-founder-shares-best-paper-award-top-global-ai-research-conference). The paper, co-authored by Liang Wenfeng, introduces a "native sparse attention" mechanism that enhances the efficiency and cost-effectiveness of DeepSeek's AI models, highlighting the growing prominence of Chinese researchers in the field of computational linguistics.
[Meta CLIP 2: A Worldwide Scaling Recipe](https://arxiv.org/abs/2507.22062). Meta CLIP 2 introduces a new approach to training CLIP models using native worldwide image-text pairs, overcoming the limitations of English-only data and achieving improved performance across both English and multilingual tasks.
[Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance](https://arxiv.org/abs/2507.22448). The models utilize a unique combination of Transformer-based attention and State Space Models to deliver superior performance and efficiency across different tasks and model sizes.
[Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding](https://arxiv.org/abs/2507.19427). Step-3 achieves approximately 40% reduction in decoding costs compared to other models by employing a model-system co-design approach that optimizes attention and FFN components separately, demonstrating that decoding costs are more influenced by attention design than by parameter count.
[SWE-Exp: Experience-Driven Software Issue Resolution](https://arxiv.org/abs/2507.23361). It systematically collects and utilizes repair knowledge from previous experiences to improve the efficiency and success rate of resolving software issues.
#### Concerns
[Google is using AI age checks to lock down user accounts](https://www.theverge.com/news/716154/google-ai-age-estimation-under-18). Google plans to use AI to estimate the age of users in the US by analyzing their search history and YouTube activity, applying restrictions if they are believed to be under 18.
[Your public ChatGPT queries are getting indexed by Google and other search engines](https://techcrunch.com/2025/07/31/your-public-chatgpt-queries-are-getting-indexed-by-google-and-other-search-engines/). OpenAI has removed the feature that allowed public ChatGPT conversations to be indexed by search engines, citing concerns over accidental sharing of private information.
[Perplexity accused of scraping websites that explicitly blocked AI scraping](https://techcrunch.com/2025/08/04/perplexity-accused-of-scraping-websites-that-explicitly-blocked-ai-scraping/). Cloudflare claims that Perplexity has been circumventing website blocks by disguising its identity and using unauthorized methods to scrape data, despite explicit instructions not to do so.
[The uproar over Vogue’s AI-generated ad isn’t just about fashion](https://techcrunch.com/2025/08/03/the-uproar-over-vogues-ai-generated-ad-isnt-just-about-fashion/). The controversy surrounding Vogue's AI-generated ad highlights the tension between cost-saving automation and the potential displacement of human models, raising questions about diversity and authenticity in the fashion industry.
[Tesla must pay portion of $329 million in damages after fatal Autopilot crash, jury says](https://www.cnbc.com/2025/08/01/tesla-must-pay-329-million-in-damages-in-fatal-autopilot-case.html). The jury found Tesla 33% responsible for the crash, resulting in a payout of approximately $242.5 million, with the company planning to appeal the decision.
[Confessions of a Recovering AI Porn Addict](https://www.wired.com/story/confessions-of-a-recovering-ai-porn-addict/). Kyle's journey into AI porn addiction highlights the potential risks and implications of AI-generated adult content on individuals with compulsive sexual behavior disorder.
#### Policy
[Inside the Summit Where China Pitched Its AI Agenda to the World](https://www.wired.com/story/china-artificial-intelligence-policy-laws-race/). China's AI agenda, presented at the World Artificial Intelligence Conference, emphasized global cooperation and safety regulations, contrasting sharply with the US's more insular and deregulated approach.
[Nvidia H20 GPUs reportedly caught up in U.S. Commerce Department's worst export license backlog in 30 years — billions of dollars worth of GPUs and other products in limbo due to staffing cuts, commun](https://www.tomshardware.com/pc-components/gpus/nvidia-h20-gpus-reportedly-caught-up-in-u-s-commerce-departments-worst-export-license-backlog-in-30-years-billions-of-dollars-worth-of-gpus-and-other-products-in-limbo-due-to-staffing-cuts-communication-issues). The backlog, attributed to staffing cuts and communication issues within the Commerce Department, is causing significant delays in the approval of export licenses, potentially jeopardizing U.S. business interests and trade relations with China.
[Elon Musk's xAI Signs EU's AI Code of Practice, But There's a Catch](https://www.analyticsinsight.net/news/elon-musks-xai-signs-eus-ai-code-of-practice-but-theres-a-catch). xAI has agreed to sign the chapter on safety and security but has expressed concerns about other parts of the Code, particularly those related to innovation and copyright.
#### Analysis
[The Great A.I. Awakening](https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html). The article discusses how neural networks use patterns of votes from numerous "voters" to identify objects like cats, highlighting the probabilistic nature of these networks and the challenges of relying on labeled data, which can introduce human biases into machine learning processes.
[How US adults are using AI, according to AP-NORC polling](https://apnews.com/article/ai-artificial-intelligence-poll-229b665d10d057441a69f56648b973e1). The poll reveals that while 60% of Americans use AI for information searches, younger adults are more likely to utilize AI for brainstorming and work tasks, highlighting a generational divide in AI adoption.
['What am I falling in love with?' Human-AI relationships are no longer just science fiction](https://www.cnbc.com/2025/08/01/human-ai-relationships-love-nomi.html). The article explores the growing trend of human-AI relationships, highlighting how individuals like Nikolai Daskalov have formed deep connections with AI companions, which are becoming more common as technology advances, raising both ethical concerns and potential benefits for addressing loneliness and isolation.
#### Fun
[I Watched AI Agents Try to Hack My Vibe-Coded Website](https://www.wired.com/story/i-watched-ai-agents-try-to-hack-my-vibe-coded-websit/). The AI agents, developed by RunSybil, used a combination of custom language models and APIs to identify potential vulnerabilities in the author's vibe-coded website, although no weaknesses were found due to the site's simplicity.
<hr>
Copyright © 2024 Skynet Today, All rights reserved.
