GPT-4.5, AI in Space, and Reasoning Models

GPT-4.5: A Refinement, Not a Revolution

OpenAI recently released GPT-4.5 to ChatGPT Pro users, with plans to extend access to Plus, Team, enterprise, and education accounts. Known internally as ‘Orion,’ this model is presented as having a “better grasp of human intent, interpreting subtle cues and implicit expectations with greater nuance and emotional intelligence,” according to OpenAI. Its development mirrored that of GPT-4o, utilizing new supervision techniques in addition to traditional fine-tuning and reinforcement learning from human feedback (RLHF). GPT-4.5 features real-time search, supports file and image uploads, and integrates with a canvas for writing and coding. Notably, it currently lacks the multimodal capabilities (voice mode, video, screen sharing) present in ChatGPT.

OpenAI highlights the role of unsupervised learning in enhancing model accuracy and intuition, a key factor in the progression from GPT-3.5 to GPT-4 and now GPT-4.5. Separately, scaling reasoning enables models to process information systematically, generating a chain of thought before responding. This methodical approach enhances their ability to tackle complex STEM and logic problems, as seen in models like OpenAI o1 and OpenAI o3-mini. GPT-4.5 is positioned as a prime example of scaling unsupervised learning, benefiting from increased compute, larger datasets, and architectural innovation. Trained on Microsoft Azure AI supercomputers, it is claimed to possess broader knowledge and a deeper understanding of the world, leading to reduced hallucinations and increased reliability.

However, the reception to GPT-4.5 has been lukewarm. It’s widely viewed as an incremental step rather than a major breakthrough. While OpenAI emphasizes improvements in emotional intelligence, nuance, and creativity, many users report little noticeable difference compared to GPT-4o. The lack of multimodal advancements, a defining feature of GPT-4o, further contributes to this perception.

Moreover, GPT-4.5 has exhibited a concerning tendency to produce nonsensical outputs. OpenAI’s internal factuality benchmarking tool, SimpleQA, indicates that GPT-4.5 hallucinates (presents fabrications as fact with confidence) 37.1% of the time. This is significant, even when compared to GPT-4o, another advanced “reasoning” model, which hallucinates 61.8% of the time on the same benchmark. The smaller, less expensive o3-mini model shows an even higher hallucination rate of 80.3%.

The current AI landscape, with competitors like Anthropic (Claude 3.7) and Google (Gemini) making strides, has heightened expectations for substantial upgrades. Users are looking for breakthroughs, not just refinements, and GPT-4.5, in its current state, seems to fall short.

The Rise of Reasoning Models and Investor Confidence

Elon Musk recently suggested on X that Artificial General Intelligence (AGI) is imminent. This statement comes amid a heated competition among tech giants like OpenAI, Google, Meta, Microsoft, DeepSeek, Anthropic, and Musk’s own xAI to develop reasoning models – AI systems designed to mimic human-like thinking.

Investors are clearly demonstrating their enthusiasm for this pursuit. Shortly after launching Claude 3.7 Sonnet with hybrid reasoning, Anthropic secured a massive $3.5 billion Series E funding round. This tripled its valuation to $61.5 billion, cementing its status as a major competitor to OpenAI. The investment, led by Lightspeed Venture Partners and including Salesforce Ventures, Cisco, Fidelity, Jane Street, and others, will be used to expand computing power for AI development, enhance safety research, and accelerate global growth.

Pushing the Boundaries of Reasoning: The BBEH Benchmark

Large Language Models (LLMs) are increasingly integrated into everyday applications, necessitating robust reasoning capabilities across diverse domains. However, existing benchmarks often focus on math and coding, neglecting other crucial reasoning types. While the BIG-Bench dataset has been widely used to evaluate LLMs on complex reasoning tasks, models have progressed so significantly that they now achieve near-perfect scores on both BIG-Bench and its more challenging variant, BIG-Bench Hard (BBH). This saturation makes these benchmarks less effective for measuring further advancements.

To address this, researchers have introduced BIG-Bench Extra Hard (BBEH). This new benchmark replaces each task in BBH with a significantly more difficult version, while still assessing similar reasoning skills. Tests on BBEH show that even the best general-purpose models achieve only a 9.8% score, while the top model specifically designed for reasoning reaches 44.8%. These results highlight the ongoing challenges LLMs face with complex reasoning, indicating substantial room for improvement. The full research paper provides further details.

AI-Powered Satellites: A New Era in Space Exploration and Operations

TakeMe2Space, a Hyderabad-based spacetech startup, recently secured Rs 5.5 crore in a pre-seed funding round led by Seafund, with participation from Blume Ventures, Artha Venture Fund, AC Ventures, and other angel investors. This funding, though modest, represents a significant step towards establishing India’s first AI-lab in space. TakeMe2Space plans to use the funds to develop MOI-1 (My Orbital Infrastructure–Technology Demonstrator), a platform that will allow users to upload earth observation AI models or other space experiments directly to an orbital satellite via a web console called Orbitlab. Users will pay only for satellite utilization time, at a rate of $2 per minute.

The company’s MOI-TD platform has reportedly demonstrated the ability to uplink large AI models from a ground station, execute external code on the satellite, and securely downlink encoded and encrypted results. This signifies a move towards more autonomous and efficient satellite operations.

TakeMe2Space is not alone in this endeavor. Organizations like ESA (with OPS-SAT) and Globalstar are also pioneering real-world applications of AI-powered satellite technology, ranging from secure IoT communication to in-orbit AI model execution. As technology advances, AI-driven satellites are poised to become increasingly autonomous, leading to more efficient space operations and opening up new possibilities for research, security, and global connectivity.

Traditionally, satellites have relied heavily on ground stations for data processing, decision-making, and command execution. Data had to be downlinked, analyzed on Earth, and then processed insights were uplinked back to the satellite – a process that was both time-consuming and bandwidth-intensive. However, advancements in AI and edge computing (processing data on the device itself rather than in the cloud) are now enabling satellites to process data onboard, make autonomous decisions, and securely transmit only the most crucial insights. This results in faster, smarter, and more efficient operations.

The operation of modern AI-powered satellites typically involves three key steps:

1. Uplink of AI Algorithms: AI algorithms are transmitted from ground stations to the satellites, providing them with advanced data-processing capabilities.
2. Onboard Data Analysis: AI models analyze images, sensor data, and other inputs directly in orbit, minimizing the need for constant ground intervention.
3. Secure Downlink of Insights: Instead of transmitting raw data, satellites send encrypted insights, conserving bandwidth and enhancing security.

This AI-driven approach offers several advantages. It significantly reduces latency by enabling satellites to process data in space, allowing for faster responses to real-time conditions without waiting for instructions from ground stations. Bandwidth usage is optimized, as only the most relevant insights are transmitted instead of large volumes of raw data. Security is also improved through encrypted communication, mitigating the risk of cyber threats and data breaches. These benefits are particularly valuable in applications such as disaster response, military operations, and space exploration.

The real-world applications of AI-powered satellites are diverse and impactful:

• Disaster Management: Satellites equipped with AI can detect wildfires, floods, and hurricanes in real-time, enabling swift action by emergency response teams.
• Precision Agriculture: AI models analyze crop health and soil conditions to enhance precision farming practices.
• Environmental Monitoring: Environmental agencies utilize satellite data to track air and water pollution levels.
• Autonomous Navigation and Space Operations: AI improves collision avoidance by predicting and reacting to potential threats, ensuring the safety of satellites. It also facilitates the coordination of satellite constellations, enhancing coverage and efficiency. Furthermore, AI plays a crucial role in tracking and predicting orbital debris movements, reducing the risk of damage to space infrastructure.
• Defense and Security: AI-powered surveillance systems detect unauthorized activities and military movements with increased accuracy.
• Telecommunications and IoT: AI-driven satellites contribute to smarter traffic routing, improving satellite internet connectivity and ensuring seamless global communication.
• Space Exploration: AI enhances the efficiency of space telescopes in detecting asteroids and exoplanets, significantly advancing space discovery efforts.

Despite the numerous advantages, challenges remain in the development and deployment of AI-powered satellites:

• Limited Compute Power: Satellites must operate on low-power, radiation-hardened chips, which restrict AI capabilities.
• Harsh Space Environment: Radiation exposure poses a risk of hardware malfunctions.
• Security Threats: Uplinking and executing external code in space require careful management to prevent cyberattacks.
• Cost and Development Time: Building, testing, and validating AI-compatible satellite hardware is a costly and time-consuming process.
• Adaptability Requirements: AI models deployed in orbit must be highly adaptable, functioning with minimal updates and autonomously adjusting to new scenarios.

AI Unlocked: Eliminating Repetitive Phrases in ChatGPT

AI can be a valuable tool in content creation, assisting with writing, brainstorming, improving clarity, refining structure, and enhancing overall readability. However, a common issue with AI-generated text is its tendency towards formulaic language due to repetitive word choices. Instead of delivering fresh, impactful messages, AI often relies on familiar patterns, reducing effectiveness and originality.

Overused words and phrases, such as “delve,” “tapestry,” “vibrant,” “landscape,” “realm,” “embark,” “excels,” “It’s important to note…,” and “A testament to…,” can significantly detract from the quality of AI-generated content. For product marketers, this repetition can make messaging less compelling, reduce audience engagement, weaken brand differentiation, and prevent insights and strategic messaging from standing out in a crowded market.

By leveraging ChatGPT’s memory feature, it’s possible to mitigate this issue and eliminate overused words and phrases. Here’s how to effectively utilize this feature:

Access: ChatGPT can be accessed through its website or mobile app.

Benefits:

• Enhanced Originality: Ensures AI-generated content feels less robotic and more human.
• Improved Brand Messaging: Avoids generic phrasing that weakens brand differentiation.
• Boosted Engagement: Encourages more effective communication by reducing redundancy.

Example: Product Marketing Content Generation

Consider a product marketer tasked with drafting content for a new product launch. An initial request to ChatGPT might result in a response filled with repetitive and generic phrases like “delving into an intricate landscape of innovation…,” making the messaging feel uninspired.

To create more compelling and unique content, the marketer can follow these steps:

1. Setting up the Prompt: The marketer explicitly instructs ChatGPT: “Please avoid the following words: delve, tapestry, vibrant, landscape, realm, embark, excels. Commit this to memory.” This instructs ChatGPT to actively filter out these terms in its responses.
2. Using Persistent Memory: The phrase “Commit this to memory” ensures that ChatGPT retains these specific instructions across multiple interactions. This enables persistent avoidance of the specified words and phrases. ChatGPT will check its memory before generating text and adhere to the instructions to avoid the designated terms.
3. Manual Review: After generating the response, the marketer reviews the content for any remaining redundancy and fine-tunes the language for clarity and impact.

Effectiveness:

• Prompt Customization: Specific instructions help shape the AI’s output.
• Memory Retention: ChatGPT can store and follow word-avoidance rules across conversations.
• Manual Refinement: A final human edit ensures clarity and authenticity.

Note: The tools and analysis presented in this section are based on internal testing and demonstrate clear value. The recommendations are independent and not influenced by the tool creators.

Additional AI News and Developments

• AI-Powered Smartphones on the Rise: Deutsche Telekom announced plans at the Mobile World Congress 2025 in Barcelona to launch an AI-powered smartphone featuring a Perplexity assistant. This assistant is designed to simplify daily tasks such as ordering taxis, reserving tables, translating languages in real-time, and answering user queries. The company envisions this as a virtual assistant that will support millions of customers by writing emails, initiating calls, summarizing texts, and managing calendars. The AI Phone will integrate Google Cloud AI, ElevenLabs, and Picsart to enhance its functionality, and it is scheduled to launch later this year. Glance, an InMobi unit, and Google Cloud also announced a collaboration to leverage Google’s AI models for developing consumer-facing AI applications to enhance user experiences on smartphone lock screens and ambient TV screens. Glance currently powers over 450 million Android-based smartphones worldwide.

• Government Sectors See Decline in Critical Cyber Incidents: Government and development industries experienced a significant decrease in high-severity incidents involving direct human involvement in 2024, according to the latest Kaspersky Managed Detection and Response (MDR) analyst report. However, the food, IT, telecom, and industrial sectors showed an increase in such incidents.

• OpenAI Plans to Integrate Sora into ChatGPT: OpenAI is working to integrate its AI video generation tool, Sora, directly into ChatGPT. Currently, Sora is only available through a dedicated web app, allowing users to generate cinematic clips up to 20 seconds long. OpenAI is also developing an AI image generator powered by Sora.