This week in AI research: OpenAI teases release of GPT-4.5

While the pace of AI news seemed to slow somewhat in late 2024, things have progressed quickly in recent months. The GenAI pioneer OpenAI is now building excitement for GPT-4.5, which would be the first model in the series to launch since 2023. (Its 4o variant dropped in May 2024.) In other news, the AI landscape continues to broaden, with developments spanning quantum computing, international partnerships, and open-source rivals. In addition, JD Vance recently called for fewer AI regulations in a talk at the AI Action Summit in Paris. 

1. OpenAI reveals plans for GPT-4.5 and beyond

OpenAI CEO Sam Altman hinted at the eventual release of GPT-4.5, the model which had the internal nickname “Orion” internally, in a post on X. Describing it as the last non-chain-of-thought model from the company, Altman said the company would unify its o-series reasoning models with its earlier “GPT” series, allowing the future models to adapt to user requests to decide how best to respond and pick the best internal approach to come up with a satisfactory answer.

When asked over X when GPT-4.5 might release, Altman said in “weeks” or months. Some industry observers had anticipated a GPT-4.5 launch in 2024 to bridge the gap before GPT-5, while the WSJ had noted toward the end of the year that the development of the successor of GPT-4 — Orion — was behind schedule and had incurred substantial cost overruns.

For context, ChatGPT based on GPT-3.5 first hit the scene in November 2022, while GPT-4 followed in March 2023.

OpenAI has not publicly disclosed detailed training costs for GPT‐3.5 or GPT‐4, but Altman has stated previously that training GPT-4 cost over $100 million. Independent estimates for training a GPT‑3–sized model (on which GPT‑3.5 is based) suggest that compute expenses alone for that model might be in the range of several million dollars.

The WSJ article last year noted that a six-month training run for a model like Orion could be in the ballpark of “half a billion dollars in computing costs alone.”

After the launch of GPT-4.5, “a top goal for us is to unify o-series models and GPT-series models by creating systems that can use all our tools, know when to think for a long time or not, and generally be useful for a very wide range of tasks,” he added. GPT-5 would draw from its o3 series. While the company recently released its o3-mini model variants, Altman clarified that it would no longer ship the full-fledged o3 as a standalone model, which it had indicated it would do in December.

2. OpenAI and SoftBank partner on AI in Japan

Source: SoftBank

OpenAI and SoftBank have formed a 50/50 joint venture called SB OpenAI Japan to develop and market AI solutions tailored for Japanese businesses, with SoftBank planning to invest $3 billion annually to deploy OpenAI’s technologies across its group companies. Affiliated firms will receive priority access to models like ChatGPT Enterprise and specialized agent products, and integration is expected in key sectors such as telecommunications, finance, and at Arm (another SoftBank subsidiary) to boost productivity. The venture includes Cristal Intelligence, a secure AI platform for major Japanese corporations, and is linked to the larger Stargate project—an AI infrastructure effort with an initial $100 billion U.S.-based investment (potentially rising to $500 billion over four years) involving Oracle and other stakeholders, rather than solely the U.S. government. While these details capture the essence of the partnership, aspects like Stargate’s funding and Cristal Intelligence’s scope are often oversimplified, underscoring the need for more nuanced reporting.

Why It Matters: The collaboration underscores growing international alliances to scale AI globally. SoftBank’s capital and OpenAI’s technology could point to broader AI adoption in Japan where AI has generally trailed other nations. An article from Nikkei notes that 84% of firms in the U.S. and 85% in China use AI, while only 47% do in Japan, citing J.P. Morgan research.

3. Bank of Montreal and other financial institutions join IBM’s quantum network

Source: BMO

Bank of Montreal (BMO) is joining IBM’s quantum network alongside 50 other financial institutions, aiming to accelerate tasks like climate-risk forecasting and fraud detection. Meanwhile, a University of Toronto and Insilico Medicine study used a hybrid of quantum computing and AI to discover potential drug candidates targeting mutant KRAS—a protein previously considered “undruggable.” Together, these developments signal how quantum computing is moving from theory to application in both finance and biomedicine.

Why It Matters: As global banks, technology giants, and academic labs invest heavily in advanced systems, the race to achieve practical quantum advantage is rapidly intensifying.

4. Researchers create open rival to OpenAI’s o1 ‘reasoning’ model for under $50

Source: TechCrunch

A joint team from Stanford and the University of Washington has unveiled “s1,” an open-source model with reasoning performance comparable to OpenAI’s o1 and DeepSeek’s R1, trained for less than $50 in cloud compute. It is available on GitHub where it currently has 1.7k stars. By distilling “thinking steps” from Google’s publicly accessible Gemini 2.0 Flash Thinking Experimental, researchers fine-tuned the lightweight Qwen base model (from Alibaba’s AI lab) on just 1,000 carefully curated questions. The result: an AI that scores highly on math and coding benchmarks using minimal resources and a simple approach to scaling its “thinking time.” The researchers outlined their methodology in an arXiv preprint.

Why It Matters: This underscores the shifting competitive landscape in AI, where distillation can rapidly clone or approximate expensive, proprietary models at a fraction of the cost. In the coming months and years, distillation will likely lead to increasingly powerful “edge” AI that can run on laptops, smartphones, providing a more distributed genAI reality, which initially was highly centralized.

4. France reveals plans for €109 AI effort

Source: Politco.eu

French President Emmanuel Macron has announced a €109 billion investment plan to boost artificial intelligence in France, positioning it as a response to the US’s recent $500 billion “Stargate” initiative announced by President Trump. The French plan, unveiled on February 10, 2025, ahead of a two-day AI summit in Paris, will be partially funded by foreign investors, including a €30 billion commitment from the UAE for a data center and €20 billion from Canadian investment firm Brookfield. Macron emphasized that the investment is proportionally equivalent to the US plan when considered on a per capita basis, stating that France needs to “be in the race” to avoid becoming dependent on others.

5. Reddit planning AI-based search rollout

Source: TechCrunch

Reddit plans to enhance its search capabilities in 2025 with an AI-powered system that can tackle complex and subjective questions, according to CEO Steve Huffman’s announcement during their Q4 2024 earnings call. The company will integrate Reddit Answers, a feature that provides curated summaries of relevant responses and threads, into its existing search functionality. While Reddit’s daily active users grew 39% year-over-year to 101.7 million (slightly below investor expectations), the platform is actively building a dedicated search team and sees this enhanced search capability as a potential driver for growth, retention, and monetization. This move is part of Reddit’s broader embrace of AI, following their previous rollout of AI-powered translation and brand insights features.

6. JD Vance Calls for “Open Regulatory Environment” on AI

Source: The Wall Street Journal (op-ed)

During a speech at the Artificial Intelligence (AI) Action Summit in Paris on Tuesday, U.S. Vice President JD Vance urged European allies not to impose heavy-handed AI regulations that could stifle innovation and inadvertently help the Chinese AI sector. Vance contrasted his stance with the Biden Administration’s prior focus on “AI safety,” arguing that too much regulation benefits large incumbents at the expense of startups. He emphasized the need for reliable energy sources to support high-compute AI tasks and warned European leaders that adopting Chinese platforms like DeepSeek poses a security risk. The speech also touched on how broad regulations can lead to a “duopoly” in which government and Big Tech control AI development, leaving less room for smaller companies and open innovation. Vance concluded that U.S. leadership in AI is not guaranteed, especially if overregulation hampers America’s ability to compete with China.

Why It Matters: Vance’s remarks highlight tensions between the push for AI oversight and fears that stringent rules could hand strategic advantages to authoritarian regimes or tech giants. By encouraging a more open regulatory environment, he suggests that nurturing smaller AI ventures fosters competition and prevents dominance by a few large players. The speech also underscores the geopolitical stakes of AI development. Competition in the field has recently heated up with the launch of DeepSeek’s R1 reasoning model.