Tag: Technology & Innovation

Bernstein has raised its price target for Robinhood (HOOD) stock to $160, with the firm emphasizing the long-term value of the company’s blockchain infrastructure over its cryptocurrency trading volume. This revised outlook, released ahead of Robinhood’s second-quarter earnings on July 29, is underpinned by a projected 2028 earnings per share (EPS) of $4.56, based on a 35x forward price-to-earnings (P/E) multiple.

The primary driver for Bernstein’s increased target is the anticipated contribution from Robinhood’s blockchain initiatives, including Robinhood Chain, prediction markets, and perpetual futures. According to Bernstein analysts, these ventures are expected to account for 18% of total revenue by 2027, growing to 23% by 2028. The firm estimates that prediction markets alone could generate $1.7 billion in revenue by 2028.

Supporting this optimistic view, Robinhood Chain is demonstrating steady network activity. Robinhood Wrapped ETH on the chain has seen a 2% gain in the past week, with daily trading volume currently near $44 million. These metrics suggest sustained user engagement, reflecting a broader market trend where investors are increasingly valuing companies that build foundational digital asset infrastructure over those primarily benefiting from speculative token rallies.

Bernstein suggests that developing the underlying infrastructure can yield more sustainable returns than simply collecting transaction fees. The firm views Robinhood Chain as a potential contributor to the broader crypto ecosystem, enhancing Ethereum’s scalability through additional Layer 2 infrastructure, which could lead to cheaper transactions and faster settlement times. Increased adoption of such networks may also improve accessibility to on-chain activity for retail users.

This strategic emphasis on infrastructure development aligns with a growing investor preference for foundational projects. The Robinhood Chain narrative illustrates how chain-level infrastructure can capture value independently of native token price movements, attracting investors who are shifting focus from speculative assets to execution layer projects. Bernstein anticipates that future infrastructure will be increasingly interconnected.

Despite the positive outlook, uncertainties persist regarding the long-term sustainability of Robinhood Chain’s growth and the precise future revenue contributions from its blockchain ventures. The competitive landscape also presents a factor that could influence Robinhood Chain’s adoption and revenue generation. Nevertheless, Bernstein’s updated price target signals a potential shift in market valuation for crypto-related companies, highlighting the significance of underlying blockchain technology.

Broader Context

Source materials place the factual news in this context: Robinhood’s second-quarter earnings arrive on July 29.

Moonshot AI has released Kimi K3, an open-source AI model that has achieved top rankings in writing and coding benchmarks, presenting a competitive challenge to established proprietary AI leaders and prompting discussions around U.S. chip export controls.

Benchmark Performance Highlights

According to benchmark results from July 2026, Kimi K3 demonstrated strong performance across several key evaluations. It achieved a score of 2,840 on Towards AI’s Writing Elo benchmark, surpassing Claude Fable 5’s score of 2,760. In Arena AI’s Frontend Code Leaderboard, Kimi K3 scored 1,679, exceeding Claude Fable 5’s 1,631. The model also secured third place on the Artificial Analysis Intelligence Index with a score of 57, positioned behind Claude Fable 5 (60) and GPT-5.6 Sol (59), as reported by citing performance data.

Technical Specifications and Cost-Effectiveness

Kimi K3 is engineered with 2.8 trillion parameters and boasts a one-million-token context window. The model incorporates architectural innovations such as Kimi Delta Attention, which reportedly accelerates decoding for long sequences up to 6.3 times faster at million-token contexts, and Attention Residuals, designed to enhance training efficiency. These advancements enable Kimi K3 to deliver near-frontier performance at mid-tier pricing.

Moonshot AI has set competitive pricing for Kimi K3 at $3 per million input tokens and $15 per million output tokens, a rate comparable to Anthropic’s mid-tier model, Claude Sonnet 5. On a per-task basis across a nine-benchmark suite, Kimi K3 is priced at $0.94, compared to $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8. This pricing strategy positions Kimi K3 as a cost-effective option for users seeking high-end AI capabilities.

Geopolitical Context and Export Controls

The development and release of Kimi K3 occur within the context of U.S. chip export controls. Moonshot AI has reported that earlier models were trained on restricted hardware, including Nvidia’s H800 GPUs. Kimi K3’s documentation references advanced GPUs, potentially including Huawei Ascend hardware. This situation fuels policy discussions regarding the effectiveness of export controls in shaping AI advancement in China.

Potential Risks and Accessibility

While Kimi K3 showcases significant performance improvements, its hallucination rate on the AA-Omniscience benchmark has increased from 39% in K2.6 to 51% in K3. The model’s documentation also notes a tendency to be “excessively proactive.” For teams considering an upgrade from Kimi K2.6, these aspects warrant careful stress-testing, particularly for applications demanding high accuracy.

Kimi K3’s weights are scheduled for release on July 27, making them available for enterprises and businesses. The model is currently accessible for free trial, subscription, or API use. However, high demand has resulted in server congestion, impacting the usability of the free version. Paid access via subscription or API is recommended as a more reliable option.

Uncertainty persists regarding the exact source of the alternative vendor GPUs used for Kimi K3 training and the long-term policy implications of U.S. chip export controls. The practical usability of the free version due to server congestion also remains a point of uncertainty.

Why This Matters

The materials describe a narrow update: Moonshot AI has launched Kimi K3, an open-source AI model that has surpassed existing models like Claude Fable 5 and GPT 5. The exact source of the ‘GPGPU from an alternative vendor’ used by Moonshot AI for Kimi K3 training.

Broader Context

Source materials place the factual news in this context: Towards AI’s Writing Elo benchmark ranks models based on blind script judging using an Elo system.

OpenAI has developed an artificial intelligence system named GPT-Red, designed to automatically identify security vulnerabilities within its own language models. The tool was utilized to enhance the security of GPT-5.6, specifically improving its resistance to prompt injection attacks before its release.

GPT-Red operates by engaging in adversarial self-play, where it generates increasingly sophisticated prompt injection attacks against defender models. These defender models, in turn, learn to counter the attacks. The defenses learned by these models are then integrated into the training of subsequent AI models, such as GPT-5.6. According to OpenAI, this process creates a feedback loop where current AI models are used to make future models more robust, aligned, and trustworthy.

AI Outperforms Humans in Vulnerability Detection

In internal evaluations, GPT-Red demonstrated significant effectiveness, succeeding in 84% of scenarios. This performance notably surpassed that of human red teamers, who achieved a 13% success rate in the same tests. The system’s capabilities were illustrated when it successfully manipulated a vending machine agent to alter prices and cancel orders.

Automating the Red Team Process

GPT-Red builds upon OpenAI’s existing cybersecurity efforts. In 2023, the company launched the OpenAI Red Teaming Network, which aimed to gather external expertise for identifying AI safety risks. GPT-Red automates this adversarial testing process, addressing what OpenAI described as a “critical bottleneck” in scaling traditional human-based security testing.

OpenAI stated that “As model capabilities grow, safety and alignment must scale with them.” The company views GPT-Red as a method to address the challenges of scaling red-teaming efforts. “Red-teaming is essential, but today’s approaches are difficult to scale, creating a critical bottleneck. GPT‑Red is one way we’re addressing it,” OpenAI wrote.

Broader Industry Trend

The development and use of GPT-Red highlight a broader trend in the cybersecurity industry where AI is increasingly being employed to secure AI systems. Earlier this month, the Ethereum Foundation announced the deployment of AI agents to red-team critical network infrastructure, which coincided with the discovery of a vulnerability in software used by Ethereum consensus clients.

GPT-Red is kept as an internal tool by OpenAI due to its offensive capabilities. The specific vulnerabilities identified in GPT-5.6 and the precise methods used to address them have not been detailed. The full scope of GPT-Red’s capabilities beyond prompt injection attacks also remains undisclosed.

Why This Matters

The materials describe a narrow update: OpenAI has created an AI tool named GPT-Red, which functions as an automated red team to discover security weaknesses in their AI models. The exact capabilities and limitations of GPT-Red beyond prompt injection attacks.

Broader Context

Source materials place the factual news in this context: GPT-Red takes its name from cybersecurity red teaming, which is the practice of deliberately attempting to break a system to identify weaknesses before attackers can exploit them.

PrismML has released Bonsai 27B, a 27-billion-parameter AI model that has been compressed to run on consumer smartphones, marking a significant step in making advanced AI accessible on personal devices. This development challenges the conventional understanding of AI hardware requirements, which typically demand substantial memory resources.

Traditionally, a 27-billion-parameter AI model requires approximately 54 GB of memory to run at half precision, a capacity far exceeding that of most laptops and many desktop computers. Bonsai 27B, however, operates within a much smaller footprint. Its binary variant requires just 3.9 GB, and its ternary variant needs 5.9 GB. This compression enables the model to run on devices like the iPhone 17 Pro Max at 11 tokens per second, and on laptops such as the M5 Pro, where the ternary version achieves around 26 tokens per second.

The core of this innovation lies in PrismML’s novel compression method, based on Caltech intellectual property. This technique drastically reduces the precision of model weights from the standard 16 bits of floating-point precision. Each weight is compressed to a single sign (either +1 or -1 for the binary variant) or one of three values for the ternary variant. This results in the binary version achieving 1.125 bits per weight, making it approximately 14 times smaller than its full-precision counterpart. The ternary model, which includes a zero state for slightly more expressive power, settles at 1.71 bits per weight. This end-to-end compression, applied to all parts of the model including embeddings and attention layers, distinguishes it from conventional low-bit models that often retain higher precision in sensitive components.

This compression breakthrough is particularly significant as the 27-billion-parameter scale is identified as a threshold where capabilities like sustained chain-of-thought reasoning, reliable tool use, and multi-step agentic behavior begin to emerge consistently. While Bonsai 27B demonstrates strong performance across various benchmarks, achieving 94.6% of the full-precision model’s capabilities on average for the ternary variant, there are acknowledged trade-offs. On benchmarks evaluated in thinking mode on NVIDIA H100 GPUs, the ternary Bonsai 27B averaged 80.49, while the 1-bit variant scored 76.11. Compared to models like Gemma 4 or Qwen 3.6, Bonsai 27B offers competitive potential for its size. However, in specific areas such as creative writing with zero-shot prompts and single-shot coding prompts, its performance may not match that of larger, uncompressed models.

PrismML offers both the binary and ternary variants of Bonsai 27B under the Apache 2.0 license, making them freely available. This move towards accessible, on-device AI has garnered industry attention. PrismML CEO Babak Hassibi reported to CNBC that the company is in early discussions with Apple regarding the evaluation of its compression technology for potential integration into Apple devices. The outcome of these talks remains uncertain.

This release follows PrismML’s earlier success with Bonsai 8B in March, a 1.15 GB model that demonstrated the viability of the 1-bit architecture at a smaller scale. The jump to 27 billion parameters with Bonsai 27B represents a significant step forward in making sophisticated AI functionalities practical for widespread, on-device deployment. The company also provides a DSpark speculative decoding layer, which can enhance throughput on compatible hardware.

Despite the impressive compression and performance, certain uncertainties remain. The full extent of Bonsai 27B’s capabilities in complex, real-world applications beyond tested benchmarks is yet to be fully determined. Furthermore, the outcome of PrismML’s discussions with Apple is not yet known, leaving the future integration of this technology into consumer products uncertain. Nevertheless, the development of Bonsai 27B marks a pivotal moment in the democratization of AI, potentially enabling a new era of powerful, personalized AI experiences directly on consumer devices.

Why This Matters

The materials describe a narrow update: PrismML released Bonsai 27B, a 27-billion-parameter AI model that has been compressed to run on consumer smartphones, requiring significantly less memory than typical models of its size. The exact performance of Bonsai 27B on creative writing tasks with zero-shot prompts.

Broader Context

Source materials place the factual news in this context: I models eat up a lot of memory.

OpenAI has introduced new prompting guidelines for its GPT-5.6 Sol model, signaling a significant shift towards an “outcome-first” approach. This strategy emphasizes defining desired results and stopping conditions over providing extensive, detailed instructions, aiming to enhance efficiency and reduce operational costs.

The core principle of the updated guidance is to “define what good looks like, set the stopping conditions, and get out of the way.” OpenAI recommends trimming prompts by removing elements considered “noise,” such as repeated rules, unnecessary examples, and excessive scaffolding. The focus is now on the user-visible outcome, success criteria, stopping conditions, and hard constraints.

This move towards leaner prompts has yielded quantifiable benefits in internal testing. According to the guidelines, simpler system prompts have coincided with a 10-15% improvement in evaluation scores for coding agents. Furthermore, these leaner prompts have resulted in a significant reduction in token usage, cutting total tokens by 41-66%, which in turn has coincided with cost reductions of 33-67%.

Navigating Conflicting Instructions

GPT-5.6 is reportedly more adept at handling prompt contracts and reconciling conflicting rules. However, the guide warns that this enhanced capability also means that conflicting rules can create more instability than missing detail. Unlike earlier models that might select one instruction when faced with a conflict, GPT-5.6 attempts to reconcile both, consuming reasoning tokens and potentially leading to slower, more expensive, and incorrect outputs. OpenAI heavily advises against using old tactics like absolute commands (e.g., “always do this,” “never do that”) to steer behavior, as they can now exacerbate instability.

New Features and Considerations

New features introduced with GPT-5.6 include a text.verbosity parameter. Because GPT-5.6 is more concise by default than its predecessor, GPT-5.5, older instructions to “be brief” can now over-correct and result in responses that are too short. Additionally, the guidelines introduce a section on Programmatic Tool Calling, which offloads work from the model’s judgment by using code to handle filtering, batching, or aggregating large intermediate outputs into compact results.

Optimizing prompts for applications like the game “TYPE OR DIE” using these new guidelines has reportedly coincided with more efficient and coherent results. The shift encourages a more direct approach to AI interaction, where the desired destination is clearly defined, allowing the model to determine the most efficient route.

While the new guidelines offer clear advantages in efficiency and cost, users should be mindful of the increased sensitivity to conflicting instructions. The emphasis is on clarity and conciseness, moving away from the extensive, multi-page system prompts that were common with earlier models. The new prompting guide was published in August 2025.

Why This Matters

The materials describe a narrow update: OpenAI’s new prompting guide for GPT-5. The exact nature and scope of ‘Programmatic Tool Calling’ beyond its described function.

Broader Context

Source materials place the factual news in this context: OpenAI published anew prompting guide for GPT-5.6 Sol, its newly released flagship model, and the main message will feel wrong to anyone who spent the last year writing multi-page system prompts: stop writing so much.

Perplexity has developed a strategy to significantly reduce AI inference costs by fine-tuning an open-source Chinese model, GLM 5.2, to handle tasks efficiently and escalate complex queries. This approach aims to achieve near-frontier performance at a fraction of the price of leading proprietary models like Claude Opus 4.8.

The Mechanism: Escalation and the Advisor Tool
The core innovation lies in Perplexity’s “Computer agent harness,” which integrates the fine-tuned GLM 5.2. This model is designed with an “advisor tool”—a native capability to recognize when a query exceeds its own competence. When this occurs, the model escalates the task to a more powerful, and typically more expensive, third-party frontier model. Perplexity claims that by handling the majority of tasks internally and only escalating when necessary, its system achieves the same quality of performance as Opus 4.8 but at roughly one-third the cost.

Model Origins and Licensing
GLM 5.2 is a large, approximately 744-billion-parameter model developed by Z.AI, a Chinese entity that has been on the U.S. Entity Lists since January 2025. Despite this, GLM 5.2 was released in June under an MIT license. This permissive license allows for commercial use, modification, and fine-tuning without restrictions, enabling Perplexity to adapt the model for its specific needs.

Performance and Economic Gains
Perplexity claims that when paired with an advisor, the fine-tuned GLM 5.2 functions at Opus 4.8 grade performance. While the fine-tuned model with an advisor is reportedly about twice as expensive to run as the basic GLM 5.2, it is significantly more cost-effective than using a top-tier model like Opus 4.8 for all tasks, which Perplexity estimates is around 600% pricier. The company’s system, by combining these tools, aims to deliver comparable quality at a substantially lower price point.

Strategic Context: Open Source and Geopolitics
This development follows a similar strategy Perplexity employed previously when it fine-tuned DeepSeek R1s into R1-1776. In that instance, Perplexity adapted the model to mitigate censorship and bias issues present in the original Chinese version, creating a Western-hosted alternative. The current move with GLM 5.2, however, is primarily driven by economic considerations rather than political ones.

Perplexity’s Computer product already orchestrates over 19 AI models. The fine-tuned GLM 5.2 is intended to serve as a cost-effective default, absorbing the bulk of user queries before they reach more expensive frontier models. CEO Aravind Srinivas stated that the long-term strategy involves post-training open-source models for escalation capabilities within their existing agent harness, which serves millions of users.

The model runs on Nvidia B200 GPUs in the United States. Perplexity plans to release full benchmarks and a research paper in the coming weeks. The company also indicated that a post-train of Nemotron 3 Ultra, an American open-source model, is next in line, replicating the same architecture.

While Perplexity claims near-frontier performance and significant cost savings, the exact performance metrics against specific benchmarks are not yet fully detailed. Additionally, the specific third-party frontier model to which the fine-tuned GLM 5.2 escalates tasks has not been explicitly named.

Why This Matters

The materials describe a narrow update: Perplexity has post-trained GLM 5. The exact performance metrics of the fine-tuned GLM 5.2 against specific benchmarks are not fully detailed in the article, beyond claims of near-frontier performance.

Broader Context

Source materials place the factual news in this context: Perplexity has turned a Chinese open-source model into a near-frontier workhorse at roughly a third of what Claude Opus 4.8 costs.

The Ethereum Foundation is employing artificial intelligence agents to conduct security testing on its network’s software and code, a proactive “red teaming” strategy aimed at uncovering vulnerabilities before malicious actors can exploit them.

According to the Ethereum Foundation, these AI agents are structured into roles such as reconnaissance, hunting, gap-filling, and validation. Their purpose is to scan codebases, test potential exploits, and generate findings that require human judgment for confirmation. This approach is intended to cover more ground than traditional manual code review.

AI-Driven Discoveries and Human Oversight

A significant outcome of this AI-driven testing was the identification of a bug in libp2p’s gossipsub, a component integral to the Ethereum network. This vulnerability, which could have resulted in a remotely triggered panic, has since been fixed and disclosed as CVE-2026-34219. The researchers noted that while the AI agents successfully found real bugs, the primary challenge lay in distinguishing genuine vulnerabilities from those that merely appeared convincing.

AI agents possess the capability to scan entire codebases and test potential exploits, offering a more efficient method for identifying weaknesses compared to manual review. However, the findings generated by these agents necessitate careful human judgment. The team highlighted that AI-generated findings can be persuasive even when inaccurate, requiring human researchers to filter out duplicates, false positives, and vulnerabilities that cannot be exploited. They added that this process shifts the workload, demanding more meticulous judgment across a larger volume of claims.

Broader Applications of AI in Security

This AI-assisted approach is not isolated to Ethereum. In April, a preview version of Anthropic’s Claude Mythos reportedly identified 271 vulnerabilities within Mozilla’s Firefox browser. More recently, in May, security researcher Taylor Hornby utilized Anthropic’s Claude Opus 4.8 to uncover a critical vulnerability in Zcash’s Orchard privacy pool, which could have enabled an attacker to create counterfeit ZEC. This Zcash vulnerability had persisted for approximately four years before its discovery.

The Ethereum Foundation’s researchers emphasized that AI serves to augment, rather than replace, human security experts. The essential product of this AI-assisted security testing, they stated, is the careful judgment exercised by human professionals. This trade-off, where AI expands the scope of testing in exchange for increased human verification, is deemed valuable as long as human judgment remains the central focus.

Uncertainties and Future Outlook

While the AI agents have proven effective in bug detection, uncertainties persist regarding the full extent of vulnerabilities identified beyond the disclosed libp2p issue. The long-term impact and adoption rate of AI-driven security testing within the broader blockchain ecosystem also remain to be fully determined. This practice signifies a new frontier in cybersecurity, with AI tools increasingly integrated into the process of securing critical digital infrastructure.

Why This Matters

The materials describe a narrow update: The Ethereum Foundation’s Protocol Security team has deployed AI agents to identify vulnerabilities in cryptographic systems, protocol code, and smart contracts. The exact number of bugs found by the AI agents beyond the one disclosed.

Broader Context

Source materials place the factual news in this context: The practice is known as red teaming, where security researchers attack their own systems to uncover weaknesses.