AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

China’s Apex Intelligence has released a 35B research model four months after its founding, according to a headline-only report. Chen Yongchao is quoted in the headline as saying AI self-improvement is stuck on “taste,” but details about the model, the comment and the company’s plans are unavailable.

China’s Apex Intelligence has released a 35B research model about four months after the company was founded, according to a report headline. The headline also attributes to Chen Yongchao the view that AI self-improvement is stuck on “taste,” though the available report does not provide his full remarks or explain what the release means by taste.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

The report identifies the new system as a 35B research model, but the figure is not defined: it could refer to a parameter count, another model-size measure or a name used in the announcement. The available information does not include a model card, technical paper, benchmark results, access terms or details of the training process. Those specifics are needed to compare the model with other systems or assess what it can do.

The headline places the release at roughly four months after Apex Intelligence’s founding. It does not provide an exact founding date, launch date or account of the company’s development during that period. It also does not say whether the model is publicly downloadable, available through an API, limited to research partners or presented as an internal research result.

Chen Yongchao’s comment is summarized as a claim that AI self-improvement is stuck on “taste.” No direct quotation beyond that phrase, job title or explanation of his argument is included in the available material. It is consequently not possible to determine whether he was referring to the ability to judge outputs, set research priorities, select useful changes, or another part of model development.

At a glance
announcementWhen: Reported roughly four months after Apex…
The developmentApex Intelligence has released a 35B research model about four months after its founding, while Chen Yongchao has raised a concern about taste in AI self-improvement.
Why Does AI Self-Improvement Lack “Taste”?

AI Research • Announcement Brief

Why Does AI Self-Improvement Lack “Taste”?

Apex Intelligence reportedly releases a 35B research model about four months after its founding. Chen Yongchao’s headline-level remark raises a separate question: how can AI judge which improvements are worth making?

35BModel designation; measure undefined
~4 mo.After company founding, per headline
1Reported model release
0Technical details in available item

01 / What is known

A release—and a reported concern

The headline connects two topics in its framing, but does not establish that the model tackles self-improvement.

Announcement

A 35B research model

The report headline says China-based Apex Intelligence released a model described as “35B.” It does not define what the figure measures.

Timing

About four months

The release is framed as occurring roughly four months after the company’s founding. Exact dates and a development history are not provided.

Attributed view

“Taste” as a bottleneck

Chen Yongchao is summarized as saying AI self-improvement is stuck on taste. The available item gives no fuller explanation.

02 / Why taste matters

Generating ideas is only half the task

Improvement also depends on selecting changes that hold up under evidence.

An AI system might propose code, experiments, or training adjustments. Deciding which proposal is useful requires criteria for quality—and evidence that those judgments work beyond a narrow test.

“Taste” could refer to judging outputs, choosing research priorities, selecting useful changes, or something else. Without Chen’s full remarks, the phrase remains a reported concern, not a technical finding.

A model release does not by itself show that the model can improve itself.

What the report lets readers assess

Announcement factsLimited
Technical evidenceNot supplied

Illustrative evidence gap, not a performance score. No benchmarks, model card, or technical paper are included in the available item.

03 / What remains open

Documentation will define the release

These details are needed to compare the model and understand the claim.

Model

What does 35B mean?

The designation could refer to a parameter count, another size measure, or an announcement name. The report does not say.

Evidence

What can it do?

Architecture, training process, benchmarks, intended use, and comparisons with other systems are unavailable.

Access

Who can use it?

Download availability, API access, research-partner access, or internal-only status are not specified.

04 / Evidence trail

From headline to a grounded assessment

Each next step would turn a brief announcement into something readers can evaluate.

01

Headline

Model release and “taste” remark are reported.

02

Technical record

Define 35B, training, intended use, and access.

03

Evaluation

Publish results and explain how quality is judged.

04

Interpretation

Clarify Chen’s claim and any link to this model.

Reader guide

Key questions

Keep the reported release separate from what remains unverified.

What has Apex Intelligence announced?

The report headline says it released a 35B research model. Technical documentation and performance results are not included.

What does “35B” mean?

The available information does not define the figure, so a specific model-size measure cannot be assumed.

What does Chen mean by “taste”?

The headline summarizes his view but does not explain whether he means output judgment, research direction, or another capability.

Can this model improve itself?

The report provides no evidence that it can, or that the release addresses the concern attributed to Chen.

Why Taste Matters to Model Development

The announcement brings two related questions into focus: what Apex Intelligence’s 35B research model can demonstrate, and how researchers judge whether AI systems can improve their own work. A model release is a concrete development, but its research importance cannot be measured from the headline alone. Evaluations, technical documentation and a clear account of the system’s intended use would help establish whether it contributes a new capability or marks an early research step.

The “taste” remark points to a broader challenge in AI self-improvement. Producing candidate changes is not the same as deciding which ones are useful. A system that proposes code, experiments or training adjustments still needs criteria for judging quality and evidence that those judgments hold up beyond a narrow test. Without Chen’s fuller explanation, the phrase should be treated as a reported concern rather than a technical finding about this model.

For readers, the distinction matters because claims about models improving themselves can sound more definitive than the evidence warrants. The release and the attributed comment are separate facts; the available report does not establish that the 35B model performs self-improvement, or that it was designed to address the issue Chen described.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Cypress Green

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Cypress Green

As an affiliate, we earn on qualifying purchases.

A Four-Month-Old Company’s Release

The report presents Apex Intelligence as a China-based company founded about four months before the model’s release. That short interval is part of the announcement’s framing, but no information is provided about the company’s founders, funding, research team, previous systems or institutional relationships. The limited details do not support a broader account of its position in the AI industry.

Similarly, “research model” signals a stated research orientation, not a complete description of availability or performance. Research models may be shared in different ways and evaluated for different purposes. Here, the report supplies no release notes or documentation to clarify the label, and it gives no comparison with existing models. The immediate context is therefore the release itself and Chen Yongchao’s reported observation—not a verified change in the state of AI self-improvement.

“AI self-improvement is stuck on “taste.””

— Chen Yongchao, as summarized in the report headline

Model Details Still Unavailable

The available report is a headline-only item, so key details cannot be confirmed from it. It does not specify what “35B” measures, describe the model’s architecture or training data, provide benchmarks, or state how researchers can access the system. The exact release date and the company’s founding date are also absent.

It is also unclear what Chen Yongchao meant by “taste,” what experience or role informs his view, and whether the remark was made in connection with the launch. No evidence is supplied that the model can improve itself, or that it addresses the judgment problem described in the headline. These points should not be inferred from the timing of the release or the wording of the report.

Documentation Will Define the Release

The next useful developments would be technical documentation from Apex Intelligence, including a clear definition of the 35B designation, evaluation results, intended research use and access arrangements. A fuller account of the launch would also establish its date and explain what the company considers the model’s contribution.

More detail from Chen Yongchao would clarify the “taste” argument: what capability he believes is missing, how that gap might be evaluated, and whether he sees it as a limitation of current systems or a barrier to particular self-improvement methods. Until such information is available, the confirmed development is the reported model release; its performance and connection to the broader claim remain unverified.

Source: rss

Key Questions

What has Apex Intelligence announced?

The report headline says the China-based company released a 35B research model. It does not provide technical documentation or performance results.

What does “35B” mean?

The available information does not define the figure. It should not be assumed to refer to a particular model-size measure without confirmation.

What does Chen Yongchao mean by “taste”?

The headline summarizes his view but gives no explanation. It is unclear whether he means judging output quality, selecting research directions or another capability.

Can the model improve itself?

The report does not say that the model performs self-improvement or provide evidence that it does. The release and Chen’s reported comment are not shown to be directly connected.

Source: news.google.com

You May Also Like

AI Lab Week In Review: 10 Moves From Seven Leading Labs

A Nov. 2–8 roundup tracks 10 AI lab developments in agent memory, scientific research, model access, video tools and usage limits.

AI Uncovered: The Intelligence That’s Changing the Game

AIThis post was created with the assistance of artificial intelligence (AI).Artificial intelligence…

Kolibri Is An Open-weight LLM From Aleph Alpha For German And English

Aleph Alpha released Kolibri, a German-English mixture-of-experts model under Apache 2.0, with weights published on Hugging Face.

Why Grok Users Reach Their Limits Faster, According To Musk

A report says Elon Musk explained why some Grok users reach limits faster, but the available details do not show his explanation or confirm a service-wide change.