Vinod Khosla’s 12 Predictions About The Future

Silicon Valley Venture Capitalist Shares Insights on Technology and Society

Vinod Khosla, a respected Silicon Valley venture capitalist, shares 12 predictions for the future from 2035 to 2049. His predictions encompass various aspects of technology and society, including the increasing accessibility of expertise, the proliferation of robots, AI's influence on entertainment and information, and advancements in personalized medicine and food production. Khosla anticipates transitioning to autonomous transit systems, clean energy solutions, and discovering more natural resources than consumed while acknowledging potential obstacles that may delay these developments. Despite the challenges, Khosla maintains an optimistic outlook on the abundant possibilities driven by technological progress.

Sponsor

Pinecone’s serverless vector database helps you deliver remarkable GenAI applications faster at up to 50x lower cost. Learn more about Pinecone serverless here.

• AI is turning into something new. In this TED talk, Mustafa Suleyman discusses the evolution of AI from a far-fetched concept to a critical technology that surpasses human capabilities in various fields. The advancement of AI has raised questions about its potential to solve major societal issues and its impact on employment and warfare. Suleyman proposes viewing AI as a "digital species" to understand its autonomous and dynamic nature better, and emphasizes the importance of developing AI with ethical considerations and safety guidelines in mind to ensure it benefits humanity without unintended negative consequences.

  

• It’s the End of the Web as We Know It - The proliferation of AI poses a significant threat to the online ecosystem by disrupting the traditional writer-reader dynamic. Search engines, like Google, utilize algorithms to filter vast amounts of web content, a process that gave rise to search engine optimization (SEO) tactics aimed at gaming these algorithms. However, generative AI, such as large language models (LLMs), could further complicate this landscape by digesting and regenerating web content, eliminating the need for human creators and potentially leading to the rise of large-language-model optimization (LLMO). This shift undermines creators' connection with their audience and their ability to earn revenue. Challenges like AI-generated misinformation and loss of quality content could degrade the value of the internet. The authors advocate for solutions prioritizing human-centric knowledge creation and nurturing creative communities on the web.

• The future of AI gadgets is just phones - The article criticizes the technology of the Humane AI Pin, asserting it uses outdated components and offers an overhyped user experience. Seeking alternatives, the author experiments with transforming various smartphones into makeshift AI wearables, encountering practical limitations such as software incompatibilities and awkward voice assistant interactions. The journey includes a futile attempt with a folding phone and progresses to more successful trials using a Pixel phone, Pixel Watch, and particularly wireless earbuds, which allow hands-free interaction with an AI assistant. The experience underscores the belief that smartphones and earbuds, with their advanced features and user acceptance, are likely to dominate the future of AI wearables, diminishing the case for dedicated AI gadgets.

  

• Here come the anti-woke AIs - Artificial intelligence (AI) development raises questions about whose values should be embodied in these systems. Companies like OpenAI, Microsoft, and Google carefully tune their AIs to avoid sensitive topics, leading to accusations of bias. A grassroots effort to create AIs with few or no guardrails reflects anyone's values. Open-source AI companies like Mistral and Alibaba release models with different philosophies. As AIs become more powerful, their values become increasingly important, as they can perform actions on our behalf without guardrails.

• To Unlock AI Spending, Microsoft, OpenAI, and Google Prep ‘Agents’ - AI providers like Microsoft, OpenAI, and Google are developing new features for conversational AI to handle complex tasks with minimal guidance from customers. Microsoft plans to announce new agent capabilities for its apps, powered by OpenAI's technology, which can automate multiple actions such as creating, sending, and tracking client invoices based on order history or rewriting an application's code in a different language and verifying it works as intended. These features belong to a class of AI software known as agents, bots that can work toward a goal with minimal guidance from people. Companies are looking to turn the excitement of ChatGPT into recurring revenue by offering software that automates harder tasks. Customers say such capabilities will be necessary to unlock more spending on AI.

• Meta Integrates Google & Bing Search Results Into AI Assistant - Meta's AI assistant, now upgraded with Llama 3, integrates Google and Bing search results, offering a seamless real-time search experience within Meta's suite of apps such as Facebook, Instagram, WhatsApp, and Messenger. Users can perform searches without platform switching, using a conversational interface. This development could reshape SEO strategies to prioritize conversational queries and AI assistant optimization. Despite testing limitations, the integration showcases Meta's investment in a more integrated AI-driven search landscape, extending to desktop via the meta.ai website and future plans for VR and smart glasses integration.

  

• Microsoft Hires Ex-Meta Platforms Exec to Boost Data Centers, OpenAI Supercomputers - Jason Taylor, a former Meta Platforms executive, is joining Microsoft in a role overseeing data center expansion and server-chip development. Microsoft's data center spending is increasing to around $50 billion this year, with plans for AI and cloud computing infrastructure. Taylor is expected to take over some of the duties of retiring Microsoft executive Brian Harry and contribute to discussions on AI supercomputers with OpenAI. Taylor has a background in infrastructure capital spending and AI infrastructure, with experience at Meta and as the chairman of the Open Compute Project's board.

• AI needs so much electricity that tech companies are getting into the energy business - As tech firms increasingly turn to artificial intelligence (AI), they're ramping up investment in sustainable energy solutions to meet the high power demands of AI technologies. OpenAI CEO Sam Altman invested in Exowatt, a solar-powered data center firm. Meanwhile, Microsoft is seeking approval for nuclear energy to support AI endeavors. Amazon, with interests in AI firm Anthropic, has funded over 100 renewable energy projects, maintaining its position as the world's largest corporate renewable energy buyer. However, this focus on green energy utilization also raises questions about the necessity of such high energy consumption by these companies in the first place.

• Google builds an AI model that can predict future weather catastrophes - Google has unveiled an AI-based weather forecasting model called SEEDS, which is touted to generate accurate and scalable weather predictions more cost-effectively than traditional physics-based models. SEEDS operates similarly to large language models and can rapidly create multiple weather scenario ensembles. Published in Science Advances on March 29, the model addresses the complexity of weather prediction, which is particularly critical as climate change increases the frequency of extreme weather events. Unlike current ensemble-based forecasting methods that struggle with extreme conditions due to costs and computational limitations, SEEDS uses a smaller number of "seeding forecasts" to extrapolate many ensembles. It demonstrated its efficacy by accurately predicting the 2022 European heatwave seven days in advance, a feat unachievable with fewer than 100 traditional predictions. SEEDS boasts negligible computing costs and high processing throughput, enabling more efficient and extensive weather forecasting.

  

• AI-Powered World Health Organization Chatbot is Flubbing some Answers - The World Health Organization (WHO) has introduced an AI-powered chatbot named SARAH to provide basic health information in eight languages, addressing topics such as mental health, tobacco use, and healthy eating. However, the bot's responses may not always be accurate due to its training on outdated data from OpenAI's ChatGPT 3.5, which was last updated in September 2021. SARAH is not designed to diagnose or discuss specific drugs, and it often directs users to consult with healthcare providers or WHO websites for more information. The WHO acknowledges that SARAH is a work in progress and seeks advice on how to improve it for emergency health situations. Safety and privacy concerns include the use of open-source data, potential cybersecurity risks, and the inability to track health patterns without voluntary survey data.

• Apple quietly beefs up AI division with acquisition of French startup -Apple has acquired Datakalab, a French startup that analyzes consumer behavior using machine learning and AI. This acquisition aims to enhance user experience and privacy features in Apple's products. Datakalab's technology will likely be integrated into existing Apple devices, such as the iPhone and Apple Watch. The acquisition aligns with Apple's pattern of acquiring health-related startups to strengthen its position in the healthcare industry.

• Tip line that guards against child sexual abuse images may struggle against AI - Stanford's Internet Observatory reports that the National Center for Missing and Exploited Children's CyberTipline, a key defense against child sexual abuse material (CSAM) on the internet, is under strain from rising report volumes, which reached 36.2 million in 2023. The influx of AI-generated CSAM exacerbates these challenges. Current report intake through a manual API is inefficient, with a lack of consistent data due to high turnover among content moderation staff in tech companies. The center faces staffing issues, inadequate integration with external data sources, and technological limitations, such as the inability to store sensitive data long-term due to legal constraints. Recommendations include enhancing the reporting API for completeness, increasing the center's budget for technological upgrades and expertise, and calls for a congressionally established commission to address AI-created CSAM.

• How the BBC used face swapping to anonymize interviewees - The BBC adopted AI "face swap" technology in a documentary to preserve interviewees' anonymity while maintaining storytelling quality. Traditional anonymity methods often lead to viewer bias; the BBC aimed to counteract this by transparently disclosing AI usage. By incorporating the Partnership on AI’s (PAI) Synthetic Media Framework, the BBC ensured that the disclosure was clear and integrated into the narrative. This approach served to respect the privacy of Alcoholics Anonymous members in the documentary "I’m An Alcoholic: Inside Recovery," while allowing for impactful storytelling. The BBC navigated technical challenges and editorial policy considerations, bolstering their policies on synthetic media use. The Framework provided guidance and opportunities for reflection on responsible AI use in media production, highlighting the importance of purposeful and authentic storytelling when employing such technologies.

• The Dawn Of A New Era: AI's Revolutionary Role In Accounting - The infusion of artificial intelligence (AI) into the accounting industry marks a significant evolution, shifting mundane tasks to intelligent machines and allowing accountants to concentrate on strategic endeavors. Major firms like EY, KPMG, and Deloitte have adopted AI for improved auditing, data analysis, and predictive analytics, enhancing their service quality and offering strategic business advice. This technological shift demands that accountants become tech-savvy and embrace continuous learning. While AI is streamlining operations for large firms, it also levels the playing field for smaller firms by providing access to advanced analytical tools. The movement toward AI in accounting promotes efficiency and strategic advisory roles while posing new challenges such as data security and ethical concerns. The industry is actively engaging with these challenges to maintain trust and integrity, ensuring a future where AI acts as a catalyst for innovation and responsible growth in accounting.

• USAF Test Pilot School and DARPA announce breakthrough in aerospace machine learning - The USAF Test Pilot School and DARPA have announced a breakthrough in aerospace machine learning, specifically in the development and testing of artificial intelligence algorithms for autonomous air-to-air combat. This collaboration was recognized as a finalist for the 2023 Robert J. Collier Trophy, highlighting the significant advancements in the machine-learning era within the aerospace industry. The teams conducted tests using the X-62A VISTA aircraft as part of DARPA's Air Combat Evolution (ACE) program, demonstrating successful dogfights over Edwards Air Force Base without the need for safety pilots to intervene. This breakthrough marks a transformational moment in combat aviation, with plans to continue advancing lessons learned onto future programs.

  

• HarperCollins Publishers and ElevenLabs to Bring More Stories to Life Through Audio - ElevenLabs has partnered with HarperCollins Publishers to produce audiobooks using advanced text-to-speech technology. This collaboration will focus on creating audio versions of selected backlist series books in foreign languages that would have been unlikely to receive audiobook adaptations. ElevenLabs' CEO Mati Staniszewski highlights the role of AI in expanding the availability of audiobooks and breaking down language barriers. Their user-friendly Projects tool enables rapid audiobook creation, offering creators voice customization and precise control over narration. While HarperCollins remains committed to traditional voice actor-led productions, the integration of text to speech serves as an adjunct to generate a more expansive and diverse audiobook portfolio in global markets.

• Drake Uses AI Tupac and Snoop Dogg Vocals on 'Taylor Made Freestyle,' References Taylor Swift's New Album 'The Tortured Poets Department' - Drake intensifies his hip-hop rivalry by dropping "Taylor Made Freestyle," a track featuring AI-generated voices of Tupac Shakur and Snoop Dogg addressing Kendrick Lamar and pressuring him to respond to Drake's previous diss "Push Ups." Drake taunts Lamar for waiting to retaliate due to the release of Taylor Swift's album potentially overshadowing his comeback. This back-and-forth began when Lamar called out Drake in "Like That," pushing Drake to retaliate with his own diss record. Amidst the clash, the community witnessed J. Cole briefly enter the fray against Lamar, highlighting the volatile dynamics at the top tier of the hip-hop industry.

  

Awesome Research Papers

AutoCrawler: A Progressive Understanding Web Agent for Web Crawler Generation - Web automation aims to streamline online tasks by replacing manual actions with automated processes for increased efficiency. However, traditional web automation tools like wrappers face issues with adaptability and scalability with new websites, while large language models (LLMs) struggle in open-world applications. The study introduces a novel approach that merges LLMs with web crawlers, specifically for vertical information web pages. This hybrid, named AutoCrawler, uses a two-stage process that leverages the hierarchical nature of HTML for better understanding and adaptability. It employs top-down and step-back techniques to learn from mistakes and refine HTML analysis, resulting in more accurate automated actions.

Teaching LLMs to use heterogeneous information sources - The paper discusses research on enhancing question-answering capabilities of large language models using HumanIQ. HumanIQ integrates solution processes based on human decision-making to combine information from various sources, improving the models' ability to provide accurate answers. In trials, integrating HumanIQ with the GPT-4 model cut its error rate by half and was preferred by users over baselines. The approach differs by using diverse, relevant examples for prompts and providing a more holistic integration of structured and unstructured data. The research found that correctness increases when multiple sources agree, with a 0% error rate in such cases. HumanIQ's methodology is not task-specific and requires no model fine-tuning, making it adaptable to other applications.

The Ethics of Advanced AI Assistants - The paper from Google DeepMind discusses the ethics of advanced AI assistants, emphasizing opportunities and risks. Holistic evaluations of human-AI interaction, multi-agent and societal level research, and foresight practices are crucial for managing risks. Topics covered include manipulation, anthropomorphism, trust, privacy, cooperation, equity, and access. The authors stress the importance of understanding AI assistant interactions with users, non-users, and society for responsible decision-making.

Automated Social Science: Language Models as Scientist and Subject - The paper details a method for generating and evaluating social science hypotheses using large language models (LLMs) enhanced by structural causal models (SCMs). SCMs define hypotheses and guide LLM agent creation, experimental setups, and data analysis. The system discerns verifiable causal relationships through experiments like negotiations, bail hearings, job interviews, and auctions. Results show LLMs predict causal effect directions but not magnitudes, unless informed by SCMs. An auction experiment confirmed auction theory alignment, with LLMs making improved predictions when utilizing SCMs, suggesting LLMs possess latent knowledge not immediately expressible without structural causal context.

The Open Medical-LLM Leaderboard: Benchmarking Large Language Models in Healthcare - In the rapidly advancing field of healthcare, Large Language Models (LLMs) such as GPT-3 and Med-PaLM 2 are enhancing medical applications by processing vast amounts of data to support healthcare professionals. However, the accuracy of these models is critical, as demonstrated by GPT-3's incorrect medication recommendation for a pregnant patient. The Open Medical-LLM Leaderboard serves as a standardized evaluation platform, assessing LLMs through a range of medical question-answering tasks using datasets like MedQA, MedMCQA, PubMedQA, and subsets of MMLU. This initiative not only benchmarks the models' medical expertise and reasoning but also aids in refining them to ensure safer applications in patient care. Further expansion of this platform includes integrating broader medical datasets and refining assessment metrics. Community involvement and collaboration are encouraged, as Open Life Science AI aims to foster the intersection of AI and healthcare for safer and more effective AI deployments.

Sponsor

Vultr is empowering the next generation of generative AI startups with access to the latest NVIDIA GPUs.

Try it yourself when you visit getvultr.com/forwardfutureai and use promo code "BERMAN300" for $300 off your first 30 days.