The person responsible for developing reasoning models at OpenAI. Huge implications
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
It is reported that ByteDance poached the technical director of Alibaba Tongyi Model with an 8-digit annual salary. Alibaba has previously applied for arbitration on the non-compete agreement
Sina Technology News December 6, noon news, some media reported that ByteDance poached Zhou Chang, the former technical director of Ali Tongyi Big Model, with a 4-2 job level and an 8-digit annual salary package. As of press time, ByteDance and Ali have not responded.
Public information shows that Zhou Chang joined Ali in 2017 and served as the technical director of Ali Tongyi Big Model. After Zhou Chang resigned in July this year, some media reported that Zhou Chang had joined ByteDance in August to engage in AI big model-related work. In November this year, Alibaba decided to apply for arbitration for the news that "Zhou Chang, a former employee of Tongyi Big Model, violated the non-compete agreement", which was also confirmed by Ali insiders.
According to the first financial report, "It's not just Zhou Chang who came to ByteDance, more than a dozen people in his team also jumped ship." ByteDance offered Zhou Chang a contract that was almost impossible to refuse: a 4-2 job level and an 8-digit annual salary package, which is about two levels higher and several times the salary according to Ali's job level system. ByteDance also gave the original team members who came with him job levels of 4-1 and 3-2 (equivalent to Alibaba's P10 and P9).
None of the models have a score of 100. The parameter is a test hard for AI but easy for humans. Since these tests still exist, it means it is closer to impossible so far. Only once 100% is reached do we even start talking true cost effectiveness.
Software also isn't like physical production. A fab that puts $1B into a failed 5 nm pilot plant can reuse the tools for a known 7 nm process with only depreciation loss. At worst the losses can be recouped by liquidating the assets. Spend $1B on training a model that's shit or unprofitable, that money is gone.
I mean this really looks to me like early steam age guys trying to make a robot God with a steam engine and mechanical computer brain rather than build a train.
I have to once again recite the PSA that artificial neural networks trained by backpropagation do not reason. It doesn't matter how impressive these test results are, these are fundamentally pattern matchers, not reasoning systems. They're a lot closer to autocomplete than they are to intelligence.
There are no implications beyond pumping more money into OpenAI's owners' stocks and fueling the hype bubble. And in a few months Chinese open source groups will release comparable systems and on and on we'll go. You ought to know the cycle, you came up with a "you are here" chart for SDF reactions to US chip sanctions.
It's still impossible...
If you need a mini nuclear reactor to incompletely mimic basic reasoning, you're not creating real intelligence.
Good thing Huawei slogan is Make It Possible
I do think there is nothing in laws of physics that prevent AGI.
Current hardware should suffice, maybe its the archecture that needs to be revamped
Open AI isnt training new models, o3 is just a larger o1 with more inference time compute thrown into it
It cost them more in compute cost to run the benchmark than they got out of winning the million dollar prize money
The use case for o3 will be narrow, think research institutions or strategic planning for C levels execs
Hence they were thinking about a $2000 month subscription... and at high end I could see them charging that much per query for the super important use cases
thats the very definition of hitting a scaling wall.... something new is needed to get to ASI
Closed AI releases expensive frontier models, then they also release cheaper/lower-capability models for public use, then open source models soon hit which plunge the cost with capabilities close to or equal to the closed AI frontier models optimised for public use
I estimate that Alibaba will soonish (in a few months max) come up with a new release
Test time compute is the new paradigm and I expect that significant resources this year will go on how to optimize planning and solution finding so that they can lower compute requirements.
I estimate that Alibaba will soonish (in a few months max) come up with a new release
Test time compute is the new paradigm and I expect that significant resources this year will go on how to optimize planning and solution finding so that they can lower compute requirements.
Non-compete agreements should imo be made illegal, but yeah a shame to see this specific poaching. I like Alibaba's approach on AI