Delay in Google’s Gemini 3.5 Pro AI Model Raises Concerns
Alphabet Inc.’s Google finds itself significantly behind schedule in the rollout of its much-anticipated Gemini 3.5 Pro, the company’s most robust AI model to date.
Insiders reveal that prolonged efforts to enhance its functionalities, particularly in coding, have contributed to this delay.
This situation has incited frustration among Google’s engineers, AI researchers, and managerial staff. There are palpable concerns that the company may forfeit its competitive edge as rivals such as Anthropic and OpenAI introduce models that surpass the capabilities currently expected from Gemini, as noted by ten current and former employees.
The intricate hierarchy of stakeholders involved in preparing models for release often leads to bottlenecks, complicating the integration of AI across Google’s expansive array of products, which includes Search, Maps, and YouTube.
Recently, both OpenAI and Meta Platforms Inc. launched new models that have outdistanced Google’s existing offerings in AI-driven coding.
In a bid to bolster these capabilities, Google updated the training data for Gemini late last month, but the resulting improvements were less than satisfactory, according to an insider. This news contributed to a drop in Google’s shares, which fell by as much as 3.2% last Thursday.
“We are swiftly deploying a diverse range of models while ensuring they remain cost-effective for our customers,” a spokesperson from Google conveyed.
Anticipations were high for the release of the 3.5 Pro during the company’s May developer conference. Additionally, Google is engaging with the U.S. government, which has ramped up scrutiny of advanced AI models, discussing capabilities and the safety standards applicable to the industry.
“We are actively testing 3.5 Pro, along with an upgraded Flash model, and collaborating productively with the U.S. government on model testing and broader frameworks,” the spokesperson elaborated.
Earlier this year, Anthropic experienced backlash after internal tests revealed alarming cybersecurity vulnerabilities in its recent models, leading to a temporary withdrawal of those offerings.
OpenAI, on the other hand, has taken a cautious approach, voluntarily limiting and staggering the release of its newest AI model following national security concerns and significant pressure from the previous administration.
Google’s widely-used products serve as an entry point for generative AI for everyday users, yielding increasingly intelligent responses.
However, getting leadership across various departments to align with this direction resembles the daunting task of boiling an ocean, as noted by a former employee.
Shifts in mandates and redundant efforts across multiple departments further compound the challenge of maintaining a coherent strategy.
Securing necessary resources for each initiative becomes an additional obstacle for success and market penetration.
In the wake of ChatGPT’s launch in late 2022, which sparked fears of obsolescence for Google’s search engine, the company declared a “code red” — a tactic aimed at cutting through the bureaucratic layers that often impede product development. Today, the race in AI is just the new normal for the organization, according to an employee.
Insiders report that Google co-founder Sergey Brin and others are urging for a swift response to capitalize on opportunities in AI coding, although progress is hampered by internal competition.
Various factions, including the Google Cloud division, research lab Google DeepMind, and the Android OS team, are concurrently working on AI coding tools, alongside several consumer product teams.
Opposition also arises from engineers within Google who espouse a more traditional view, arguing that all crucial code should be human-generated to uphold Google’s standards.
Initially, teams encountered restrictions using Gemini for software development due to concerns regarding proprietary code potentially contaminating the training dataset.
Although these policies have been relaxed, they limited opportunities for experimentation in AI development.
Google recently announced at its Cloud conference that 75% of its code is now AI-generated, indicating it has undergone review and meets the company’s standards for production.
The tech giant has consolidated its coding tools primarily under Google Antigravity, designed to enhance interaction with operating systems and applications through robust data, memory, and safety protocols.
To mitigate internal confusion, Chief AI Architect Koray Kavukcuoglu collaborates with the core engineering team to unify Google’s internal AI coding tools.
Furthermore, a dedicated team within DeepMind, led by research engineer Sebastian Borgeaud, has been established to focus specifically on AI coding.
While engineers are now expected to leverage AI for code generation, they often confront capacity constraints due to fierce competition for computing resources within the company.
Experts in AI assert that Gemini’s strongest advantage lies in its ability to query Google’s vast search data. However, both Anthropic and OpenAI continue to lead the charge in developing the most advanced AI models.
Google points to its unique strengths, including proficiency in handling diverse input types, such as images and videos, along with advancements in AI world models that emulate physical environments.
Frustration surrounding Google’s position in the AI landscape has fueled a noticeable exodus of talent toward Anthropic and other leading laboratories, according to former employees.
Access to Anthropic’s Claude model has been restricted to select teams engaged in high-priority research projects, limiting its availability across the organization.
While users await the release of Gemini 3.5 Pro, experiences with Gemini 3.5 Flash have been mixed. Rodrigo Davies, a product manager at the design platform Figma, reported that the company recently integrated 3.5 Flash into its “Figma agent,” an AI assistant aiding designers in generating and refining ideas. For Figma, the model strikes an ideal balance between speed and quality.
Conversely, Freddy Vega, CEO and founder of the Latin American educational technology platform Platzi, characterized 3.5 Flash as being in an awkward middle ground: more costly than Google’s previous 3.1 Flash model, yet slower and still significantly less capable than offerings from competitors.
His team has transitioned away from Google, opting instead for one of Anthropic’s mid-tier models for tasks that require a combination of speed and reasoning.
Source link: Latimes.com.
