Be a part of our each day and weekly newsletters for the newest updates and unique content material on industry-leading AI protection. Be taught Extra
There’s a brand new king on the town: Matt Shumer, co-founder and CEO of AI writing startup HyperWrite, at the moment unveiled Reflection 70B, a brand new giant language mannequin (LLM) based mostly on Meta’s open supply Llama 3.1-70B Instruct that leverages a brand new error self-correction method and boasts superior efficiency on third-party benchmarks.
As Shumer introduced in a publish on the social community X, Reflection-70B now seems to be “the world’s top open-source AI model.”
He posted the next chart exhibiting its benchmark efficiency right here:
Reflection 70B has been rigorously examined throughout a number of benchmarks, together with MMLU and HumanEval, utilizing LMSys’s LLM Decontaminator to make sure the outcomes are free from contamination. These benchmarks present Reflection persistently outperforming fashions from Meta’s Llama collection and competing head-to-head with prime business fashions.
You may strive it your self right here as a demo on a “playground” web site, however as Shumer famous on X, the announcement of the brand new king of open-source AI fashions has flooded the demo website with site visitors and his workforce is scrambling to seek out sufficient GPUs (graphics processing items, the precious chips from Nvidia and others used to coach and run most generative AI fashions) to spin as much as meet the demand.
How Reflection 70B stands aside
Shumer emphasised that Reflection 70B isn’t simply aggressive with top-tier fashions however brings distinctive capabilities to the desk, particularly, error identification and correction.
As Shumer informed VentureBeat over DM: “I’ve been thinking about this idea for months now. LLMs hallucinate, but they can’t course-correct. What would happen if you taught an LLM how to recognize and fix its own mistakes?”
Therefore the title, “Reflection” — a mannequin that may replicate on its generated textual content and assess its accuracy earlier than delivering it as outputs to the consumer.
The mannequin’s benefit lies in a method referred to as reflection tuning, which permits it to detect errors in its personal reasoning and proper them earlier than finalizing a response.
Reflection 70B introduces a number of new particular tokens for reasoning and error correction, making it simpler for customers to work together with the mannequin in a extra structured means. Throughout inference, the mannequin outputs its reasoning inside particular tags, permitting for real-time corrections if it detects a mistake.
The playground demo website consists of prompt prompts for the consumer to make use of, asking Reflection 70B what number of letter “r” cases there are within the phrase “Strawberry” and which quantity is bigger, 9.11 or 9.9, two easy issues many AI fashions — together with main proprietary ones — fail to get proper persistently. Our assessments of it had been sluggish, however Reflection 70B finally supplied the right response after 60+ seconds.
This makes the mannequin significantly helpful for duties requiring excessive accuracy, because it separates reasoning into distinct steps to enhance precision. The mannequin is accessible for obtain through the AI code repository Hugging Face, and API entry is ready to be obtainable later at the moment via GPU service supplier Hyperbolic Labs.
An much more highly effective, bigger mannequin on the best way
The discharge of Reflection 70B is simply the start of the Reflection collection. Shumer has introduced that a good bigger mannequin, Reflection 405B, shall be made obtainable subsequent week.
He additionally informed VentureBeat that HyperWrite is engaged on integrating the Reflection 70B mannequin into its main AI writing assistant product.
“We’re exploring a number of ways to integrate the model into HyperWrite — I’ll share more on this soon,” he pledged.
Reflection 405B is predicted to outperform even the highest closed-source fashions available on the market at the moment. Shumer additionally stated HyperWrite would launch a report detailing the coaching course of and benchmarks, offering insights into the improvements that energy Reflection fashions.
The underlying mannequin for Reflection 70B is constructed on Meta’s Llama 3.1 70B Instruct and makes use of the inventory Llama chat format, guaranteeing compatibility with present instruments and pipelines.
Shumer credit Glaive for enabling speedy AI mannequin coaching
A key contributor to Reflection 70B’s success is the artificial information generated by Glaive, a startup specializing within the creation of use-case-specific datasets.
Glaive’s platform permits the speedy coaching of small, extremely targeted language fashions, serving to to democratize entry to AI instruments. Based by Dutch engineer Sahil Chaudhary, Glaive focuses on fixing one of many largest bottlenecks in AI growth: the supply of high-quality, task-specific information.
Glaive’s method is to create artificial datasets tailor-made to particular wants, permitting corporations to fine-tune fashions rapidly and affordably. The corporate has already demonstrated success with smaller fashions, comparable to a 3B parameter mannequin that outperformed many bigger open-source options on duties like HumanEval. Spark Capital led a $3.5 million seed spherical for Glaive greater than a 12 months in the past, supporting Sahil’s imaginative and prescient of making a commoditized AI ecosystem the place specialist fashions may be educated simply for any activity.
By leveraging Glaive’s expertise, the Reflection workforce was capable of quickly generate high-quality artificial information to coach Reflection 70B. Shumer credit Sahil and the Glaive AI platform for accelerating the event course of, with information generated in hours moderately than weeks.
In complete, the coaching course of took three weeks, in accordance with Shumer in a direct message to VentureBeat. “We trained five iterations of the model over three weeks,” he wrote. “The dataset is entirely custom, built using Glaive’s synthetic data generation systems.”
HyperWrite is a uncommon Lengthy Island AI startup
At first look, it looks as if Reflection 70B got here from nowhere. However Shumer has been on the AI recreation for years.
He based his firm, initially referred to as Otherside AI, in 2020 alongside Jason Kuperberg. It was initially based mostly in Melville, New York, a hamlet about an hour’s drive east of New York Metropolis on Lengthy Island.
It gained traction round its signature product, HyperWrite, which began as a Chrome extension for shoppers to craft emails and responses based mostly on bullet factors, however has developed to deal with duties comparable to drafting essays, summarizing textual content, and even organizing emails. HyperWrite counted two million customers as of November 2023 and earned the co-founding duo a spot on Forbes‘ annual “30 Under 30” Listing, finally spurring Shumer and Kuperberg and their rising workforce to alter the title of the corporate to it.
HyperWrite’s newest spherical, disclosed in March 2023, noticed a $2.8 million injection from traders together with Madrona Enterprise Group. With this funding, HyperWrite has launched new AI-driven options, comparable to turning internet browsers into digital butlers that may deal with duties starting from reserving flights to discovering job candidates on LinkedIn.
Shumer notes that accuracy and security stay prime priorities for HyperWrite, particularly as they discover complicated automation duties. The platform continues to be refining its private assistant device by monitoring and making enhancements based mostly on consumer suggestions. This cautious method, just like the structured reasoning and reflection embedded in Reflection 70B, exhibits Shumer’s dedication to precision and duty in AI growth.
What’s subsequent for HyperWrite and the Reflection AI mannequin household?
Trying forward, Shumer has even larger plans for the Reflection collection. With Reflection 405B set to launch quickly, he believes it’s going to surpass the efficiency of even proprietary or closed-source LLMs comparable to OpenAI’s GPT-4o, presently the worldwide chief, by a major margin.
That’s dangerous information not just for OpenAI — which is reportedly in search of to boost a major new spherical of personal funding from the likes of Nvidia and Apple — however different closed-source mannequin suppliers comparable to Anthropic and even Microsoft.
It seems that as soon as once more within the fast-moving gen AI house, the stability of energy has shifted.
For now, the discharge of Reflection 70B marks a major milestone for open-source AI, giving builders and researchers entry to a robust device that rivals the capabilities of proprietary fashions. As AI continues to evolve, Reflection’s distinctive method to reasoning and error correction might set a brand new normal for what open-source fashions can obtain.
