While OpenAI charges $60 per million tokens for its flagship reasoning model, a Chinese startup just open-sourced an alternative that matches its performance—at 95% less cost. Meet DeepSeek-R1, the RL-trained model that’s not just competing with Silicon Valley’s AI giants, but in some cases running on consumer laptops in some configurations rather than in data centers. Meanwhile, Insilico Medicine has dosed the first patient in a Phase I trial of its AI-discovered TEAD inhibitor, which is one of “10 programs” to move from in silico into the clinic. Read on to learn more.
Insilico Medicine doses first patient in AI-discovered cancer drug trial
Source: Insilico Medicine
Insilico Medicine this week announced that the first patient was dosed in a global multicenter Phase I trial of ISM6331—its novel pan-TEAD inhibitor designed by Insilico’s generative AI chemistry engine, Chemistry42. “Just like SpaceX, we now consistently launch AI-discovered drugs into clinical trials and while many companies struggle to progress even one program into the clinic, for us another clinical trial milestone is ‘just another day in the office,’” Alex Zhavoronkov, PhD, Founder and CEO of Insilico Medicine, said in an emailed announcement.
![[An infographic from Insilico Medicine released with the news on the TEAD drug]](https://www.rdworldonline.com/wp-content/uploads/2025/01/unnamed-5.png)
[An infographic from Insilico Medicine released with the news on the TEAD inhibitor]
The news highlights how Insilico is pairing AI-driven drug design with frontier computing tools to tackle historically challenging targets. TEAD inhibitors have faced developmental hurdles as a result of off-target effects and limited efficacy, while KRAS—long considered “undruggable”—has only recently seen clinical breakthroughs. Insilico’s platform is uniquely positioned to optimize both target classes: Chemistry42 generated more than 5,000 TEAD-binding molecules before ISM6331’s selection, while the quantum-computed KRAS inhibitor exploited orbital interactions invisible to classical molecular modeling.
In parallel with the Phase I trial, the company also shared news of a recent publication in Nature Biotechnology describing the use of a real quantum computer to generate a novel KRAS inhibitor—a molecule subsequently synthesized and tested in preclinical studies.
The lowdown on Deepseek’s opensource R1 model
Source: LinkedIn / Hugging Face

R1 had “aha” moments in the training process. The image here is from the Deepseek R1 paper.
DeepSeek-R1 optimizes costs through methodological rebellion. Unlike OpenAI’s reliance on supervised fine-tuning (SFT)—a process detailed in GPT-4’s technical report—DeepSeek applied pure reinforcement learning (RL) to its base model, bypassing SFT entirely. As outlined in a Hugging Face announcement, this approach incentivized the AI to self-discover chain-of-thought reasoning through trial-and-error, yielding behaviors like self-verification and error correction absent in SFT-heavy pipelines. By contrast, OpenAI has employed a three-stage pipeline (SFT → reward modeling → PPO optimization).
OpenAI’s operational scale underscores the financial chasm between proprietary and open AI. According to Business Insider, the company spent $700,000 daily in 2023 on infrastructure alone, with 2024 projections nearing $7 billion annually for training and inference. Training its GPT-3 model in 2020 cost $4.6 million, while its unreleased GPT-5 (codenamed “Orion”) reportedly required $500 million per six-month training cycle, according to WSJ. Now consider that DeepSeek reportedly had a $12 million RL training budget for R1.
Unlike the typical approach of pretraining on massive labeled datasets and then fine-tuning with human feedback, DeepSeek-R1’s creators dropped it straight into large-scale reinforcement learning (RL). Dubbed “DeepSeek-R1-Zero,” this first version of the model learned chain-of-thought reasoning purely from trial-and-error feedback, without any supervised instruction to guide it. By pushing the model to solve ever more complex tasks, RL instilled capabilities like self-reflection and verification. The downside? The output lacked polish: DeepSeek-R1-Zero was prone to repetition, odd language mixing, and unwieldy text. To refine it, the team layered in a multi-stage process combining RL with traditional supervised fine-tuning. The resulting R1 is a system that not only matches OpenAI’s flagship on math, coding, and logic benchmarks, but also delivers output at a fraction of the cost.
Alex Zhavoronkov (Founder and CEO at Insilico Medicine) remarks:
“The AI world shrugged this week waking up to the realization that a little-known Chinese company with limited funding and computational resources dropped a completely open-source model (MIT license) on Hugging Face. The model outperforms Open AI’s frontier o1 model in benchmarks. They also explained their secret sauce in the open paper. Their secret – there is no secret. It is Reinforcement Learning.”
“Reinforcement learning is your best friend and the best RL data comes from real experiments IMHO. The most valuable data we have is the clean fully-connected data from 22 PCC nominations 10 of which went clinical.”
“Read the R1 paper by DeepSeek and embrace RL. RL is your best friend.”

[Bar chart from the Hugging Face announcement]
Trump reveals AI infrastructure blitz with $500B price tag
Source: CBS News
President Donald Trump is poised to announce a significant private sector initiative to bolster U.S. AI infrastructure, with tech giants OpenAI, SoftBank, and Oracle forming a joint venture dubbed “Stargate” that could inject up to $500 billion into AI development over the next four years, CBS News has learned. The companies plan an initial $100 billion investment, with executives including SoftBank’s Masayoshi Son, OpenAI’s Sam Altman, and Oracle’s Larry Ellison expected to join Trump at the White House on January 21, 2025, to detail the partnership.
“Together, these world-leading technology giants are announcing the formation of Stargate,” Trump said. He went on to describe it as “a new American company” that will invest in AI infrastructure in the United States. He said it would move “very rapidly, creating over 100,000 American jobs almost immediately.” An OpenAI press released noted that Stargate would eventually create “hundreds of thousands of American jobs,”
This is not the first the world has heard of Stargate. Fortune and others described a $100 billion supercomputing cluster, then described as a joint initiative between Microsoft and OpenAI, in 2024.
OpenAI enters cell reprogramming space
Source: LinkedIn, Technology Review
OpenAI is hooking up with Retro Biosciences—a longevity startup funded by OpenAI CEO Sam Altman—to create GPT-4b micro, a specialized AI model aimed at boosting the function of Yamanaka factors (proteins used to convert adult cells into stem cells). In tests, the model suggested protein redesigns that improved reprogramming efficiency by over 50%. The move is OpenAI’s first major push into biological data and a high-profile attempt to demonstrate tapping AI to make scientific discoveries. Outside scientists still await peer-reviewed validation of the results. On LinkedIn, Andrii Buvailo and Roman Kasianov of BiopharmaTrend.com highlights the news, which was also highlighted recently in Technology Review.
Trump scraps Biden’s AI executive order
Source: Reuters
President Donald Trump on Monday revoked a 2023 executive order Biden signed that required developers of AI systems posing risks to U.S. national security, the economy, public health or safety to share safety test results with the federal government under the Defense Production Act. The Biden order had directed agencies to establish testing standards and address chemical, biological, radiological, nuclear, and cybersecurity risks. The Republican Party platform characterized the order as hindering AI innovation and advocated for “AI development rooted in free speech and human flourishing.” While Trump repealed this regulatory measure, he maintained a separate recent Biden executive order focused on providing federal support for addressing energy requirements of advanced AI data centers through Defense and Energy department site leasing. The move comes as industry players like NVIDIA have expressed concerns over new Commerce Department restrictions on AI chip exports.
Samsung’s AI-focused Galaxy S25 Ultra ships February 7 for $1,300
Source: Tech Crunch
Samsung has unveiled its latest flagship smartphones at its Samsung Unpacked 2025 event, with AI taking center stage across the new Galaxy S25 series. The premium Galaxy S25 Ultra, priced at $1,300, leads the lineup with specifications including a 6.9-inch display and a quad-camera system featuring a 200-megapixel main sensor.
Key Specs
Snapdragon 8 Elite with 37% CPU and 30% GPU performance boost
50MP ultra-wide, 200MP wide, dual telephoto (3x and 5x optical zoom)
12GB RAM, up to 1TB storage
5,000mAh
OpenAI rolls out “Operator” to automate browser tasks
Source: OpenAI
OpenAI this week introduced Operator, a new AI-powered browser agent that uses computer vision and reinforcement learning to navigate websites and perform actions—like filling out forms or placing orders—on the user’s behalf. Built on a model called Computer-Using Agent (CUA), Operator is trained to “see” a website’s layout through screenshots and “interact” via simulated mouse clicks and keystrokes. Other companies such as Replit have announced agents in recent months — theirs is capable of building software applications based on user input. From the sound of it, OpenAI’s Operator will be more of a general-purpose agent capable of can tackle repetitive tasks such as reordering groceries, booking travel, and even meme creation. Operator is currently in a limited research preview, offered to U.S.-based Pro users at operator.chatgpt.com before a planned expansion to ChatGPT’s wider user base.
Even at this early stage, Operator has drawn the attention of leading online service providers. Companies like Instacart, DoorDash, Priceline, and Uber are partnering with OpenAI to optimize their user flows, while public sector groups—including the City of Stockton—plan to pilot the tech for improving municipal services access.

[OpenAI]