Send Link mailing list submissions to
[email protected]
To subscribe or unsubscribe via the World Wide Web, visit
https://mailman.anu.edu.au/mailman/listinfo/link
or, via email, send a message with subject or body 'help' to
[email protected]
You can reach the person managing the list at
[email protected]
When replying, please edit your Subject line so it is more specific
than "Re: Contents of Link digest..."
Today's Topics:
1. Re: OpenAI?s o3 Model: Breakthrough or Breakdown? (David)
2. New York Times: 'DOGE Is Building a Surveillance State'
(Stephen Loosley)
3. 'The real AI chip wars are just beginning.' (Stephen Loosley)
----------------------------------------------------------------------
Message: 1
Date: Thu, 01 May 2025 21:58:00 +1000
From: David <[email protected]>
To: [email protected]
Subject: Re: [LINK] OpenAI?s o3 Model: Breakthrough or Breakdown?
Message-ID: <2260847.h6RI2rZIcs@ulysses>
Content-Type: text/plain; charset="us-ascii"
On Monday, 28 April 2025 14:57:33 AEST Scott Howard wrote:
> For a simple case like the one you've mentioned they probably just do it in
> their head, but in a more generic sense the easy answer is to outsource that
> calculation to something designed explicitly to do that type of action - such
> as a calculator. In an agentic AI world, the role of the AI isn't to add
> those 2 numbers together, but to outsource it to an agent (such as a
> calculator, via an API) that will do it for it.\
But isn't that just moving the problem out of sight? The original issue was
that an AI system would not recognise the difference between the hard logical
proposition 2+2=4 and the softer existential (empirical) proposition "2 plus 2
usually makes 4" (using these as illustrative rather than literal examples).
Even an agenic AI system with access to the appropriate resource(s) still has
to initially recognise two distinct cases.
Then there's the issue of ultimate legal responsibility, which has to be
carried by a warm human being. If I were an Engineer designing a structure or
a Surgeon planning a procedure where safety of life was critical, I'd want to
make damned sure any AI system, agenic or not, could be relied upon to the
point where human lives and my professional reputation were not in danger. So
I imagine there would soon come a point where the human involved might as well
do the whole job themselves.
I think AI systems are very important and extremely useful in many situations,
but we need to know their constraints and how to use them.
_David Lochrin_
-------
> This seems to me to demonstrate a fundamental issue with AI machines: they
> > are still very large correlation processors but have not been developed far
> > enough to distinguish "empirical correlations" and logical rules. Thus "2
> > plus 2 usually makes 4" is an empirical correlation, but "2+2=4" in an
> > appropriate mathematical context is a logical rule. (And we understand "He
> > added 2 and 2 and got 5" isn't an appropriate context!)
>
>
> This is one of the many areas where the idea of 'agentic AI' comes in. If
> you ask a human to add two numbers, what would they do?
>
> Mix this with the newer reasoning models which are much better at working
> out the best path to come to an answer, and the types of answers you get
> from AI systems now days for this type of question is significantly better
> than it was only a few months ago. Instead of simply looking at "2 plus 2"
> and trying to guess what comes next, the recent models are able to look at
> that statement, determine it's a calculation, decide that the best way to
> solve such a statement is by using a calculator, and then call a calculator
> agent to actually do the work and get the answer. If the calculation was
> even more complex, then they might instead decide that the best option is
> to write python or R code to solve it, run that code, and then return the
> answer.
>
> Scott
> _______________________________________________
> Link mailing list
> [email protected]
> https://mailman.anu.edu.au/mailman/listinfo/link
>
------------------------------
Message: 2
Date: Thu, 01 May 2025 22:18:03 +0930
From: Stephen Loosley <[email protected]>
To: "link" <[email protected]>
Subject: [LINK] New York Times: 'DOGE Is Building a Surveillance
State'
Message-ID: <[email protected]>
Content-Type: text/plain; charset="UTF-8"
?This Is What We Were Always Scared of?: DOGE Is Building a Surveillance State
By Julia Angwin April 30, 2025
https://www.nytimes.com/2025/04/30/opinion/musk-doge-data-ai.html?unlocked_article_code=1.Dk8.lzc4.DBAR8L6qNDOp&smid=url-share
Ms. Angwin, a contributing NYTimes Opinion writer, is an investigative
journalist.
Elon Musk may be stepping back from running the so-called Department of
Government Efficiency, but his legacy there is already secured.
DOGE is assembling a sprawling domestic surveillance system for the Trump
administration ? the likes of which we have never seen in the United States.
President Trump could soon have the tools to satisfy his many grievances by
swiftly locating compromising information about his political opponents or
anyone who simply annoys him.
The administration has already declared that it plans to comb through tax
records to find the addresses of immigrants it is investigating ? a plan so
morally and legally challenged, it prompted several top I.R.S. officials to
quit in protest.
Some federal workers have been told that DOGE is using artificial intelligence
to sift through their communications to identify people who harbor anti-Musk or
-Trump sentiment (and presumably punish or fire them).
What this amounts to is a stunningly fast reversal of our long history of
siloing government data to prevent its misuse.
In their first 100 days, Mr. Musk and Mr. Trump have knocked down the barriers
that were intended to prevent them from creating dossiers on every U.S.
resident. Now they seem to be building a defining feature of many authoritarian
regimes: comprehensive files on everyone so they can punish those who protest.
?This is what we were always scared of,? said Kevin Bankston, a longtime civil
liberties lawyer and a senior adviser on A.I. governance at the Center for
Democracy & Technology, a policy and civil rights organization. ?The
infrastructure for turnkey totalitarian
Over the past 100 days, DOGE teams have grabbed personal data about U.S.
residents from dozens of federal databases and are reportedly merging it all
into a master database at the Department of Homeland Security.
This month House Democratic lawmakers reported that a whistle-blower had come
forward to reveal that the master database will combine data from such federal
agencies as the Social Security Administration, the Internal Revenue Service
and the Department of Health and Human Services.
The whistle-blower also alleged that DOGE workers are filling backpacks with
multiple laptops, each one loaded with purloined agency data.
For years, privacy advocates, including me, have obsessed about just how much
of our data Big Tech companies possess.
They know our locations, monitor our browsing and online shopping histories and
use that info to make inferences about our interests and habits.
But government records contain far more sensitive information than even the
tech giants possess ? our incomes, our bank account numbers, if we were fired,
what diseases we have, how much we gamble.
In 2009 the Georgetown law professor Paul Ohm envisioned the assemblage of a
DOGE-like amount of data and called it the ?database of ruin.? ?Almost every
person in the developed world can be linked to at least one fact in a computer
database that an adversary could use for blackmail, discrimination, harassment
or financial or identity theft,? he wrote.
We are not all the way down the rabbit hole yet. It appears that DOGE has not
yet tried to scoop up data from the intelligence agencies, such as the National
Security Agency, which collect vast amounts of communications between
foreigners ? and often catch Americans? communications in their net. (That
said, it is not encouraging that the head of the N.S.A. was recently fired,
apparently at the behest of an online influencer who is friends with the
president.)
Even so, the creation of a huge government database of personal information
about U.S. residents is dangerous and very likely against the law. In the
1960s, the Johnson administration proposed combining all of its federal
dossiers together into a new national databank. The administration said it just
wanted to eliminate duplicate records and perform statistical analysis, but the
public was outraged.
The databank was scuttled, and Congress passed the Federal Privacy Act of 1974,
which requires federal agencies to obtain consent before disclosing
individuals? data across agencies.
Of the more than 30 lawsuits that involve DOGE, several allege that its data
incursions violate the Privacy Act.
So far, courts have ruled in plaintiffs? favor in two of those cases, issuing
orders limiting DOGE?s access to data at the Social Security Administration and
Department of Treasury. Both cases are ongoing. While the orders restricted
DOGE from obtaining personally identifiable data, it remains unclear what
happens with data that has been already collected.
But the deeper problem is that the Privacy Act lacks real teeth. It did not
give judges the ability to levy meaningful fines or easily halt illegal
actions. It failed to establish an enforcement arm to investigate privacy
violations in ways that courts can?t. And since then, Congress hasn?t been able
to pass comprehensive privacy laws or create stronger enforcement mechanisms.
That makes the United States the only country in the 38-member Organization for
Economic Cooperation and Development without a data protection agency to
enforce comprehensive privacy laws. In the European Union, each country has a
dedicated data protection authority that can conduct investigations, write
rules, issue fines and even demand a halt to data processing.
Without a privacy cop on the beat, Americans can submit a Privacy Act request
to try to find out what data DOGE is holding about them or hope that judges
side with them in one of the dozens of lawsuits winding their way through
court. Still, DOGE continues going from agency to agency grabbing data.
To pick just two recent examples: Last month DOGE bullied its way into the
federal payroll records for about 276,000 federal workers, placing the
officials who objected on administrative leave, and this month a separate
whistle-blower at the National Labor Relations Board came forward with evidence
showing that after DOGE workers arrived, there was a spike in data being
siphoned out of the agency.
?In no other country could a person like Elon Musk rummage through government
databases and gather up the personal data of government employees, taxpayers
and veterans,? said Marc Rotenberg, a longtime privacy lawyer and founder of
the Center for A.I. and Digital Policy, a nonprofit research group. ?There are
many U.S. privacy laws. But they are only effective when enforced by dedicated
privacy agencies.?
We urgently need to modernize our approach to privacy by creating a federal
data protection agency with robust investigative powers.
But short of that, we still have time to stop the creation of the database of
ruin. Congress could defund DOGE or repeal Mr. Trump?s executive order
establishing it or support legislation that the Democratic senators Ed Markey
and Ron Wyden have introduced to update the Privacy Act to provide more
meaningful fines and criminal penalties.
This should be a bipartisan issue. Because once we create a database of ruin,
none of us are safe from having our information ? no matter how innocuous ?
used against us.
The Times is committed to publishing a diversity of letters to the editor. We?d
like to hear what you think about this or any of our articles. Here are some
tips. And here?s our email: [email protected].
Julia Angwin, a contributing Opinion writer and the founder of Proof News,
writes about tech policy. You can follow her on Bluesky, Twitter or Mastodon
or her personal newsletter.
Read 933 Comments ...
--
------------------------------
Message: 3
Date: Thu, 01 May 2025 23:43:39 +0930
From: Stephen Loosley <[email protected]>
To: "link" <[email protected]>
Subject: [LINK] 'The real AI chip wars are just beginning.'
Message-ID: <[email protected]>
Content-Type: text/plain; charset="UTF-8"
Huawei Ascend 910D vs Nvidia H100: A 2025 Performance Battle for AI Supremacy
News Desk May 1, 2025
https://www.reddit.com/r/technology/comments/1kc5zdl/huawei_ascend_910d_vs_nvidia_h100_a_2025/
Huawei's Ascend 910D: The Silent Challenger to Nvidia?s AI Crown ? A Deep
Global Perspective (2025)
In April 2025, Huawei's Ascend 910D emerged as a competitor to Nvidia,
previously dominating AI hardware.
Despite substantial R&D investment and a focus on specific global markets,
Huawei faces challenges such as manufacturing constraints and limited software
adoption compared to Nvidia's established ecosystem.
The competition for AI chip supremacy is intensifying.
In 2019, when the U.S. blacklisted Huawei over national security concerns, few
imagined that just six years later, Huawei would deliver a direct challenge to
Nvidia ? the unchallenged king of AI hardware.
Yet in April 2025, Huawei?s Ascend 910D has forced global chip watchers,
investors, and AI builders to take notice.
Historical Rivalry: How Huawei and Nvidia Came to Clash
Nvidia dominates AI training chips globally ? with over 80% market share as of
2024.
Huawei, initially a telecom giant, realized post-sanctions that building an
independent AI chip industry was critical for China?s technological sovereignty.
Thus began a secretive, heavily funded push into AI semiconductors, resulting
in the Ascend chip family.
How Big Are These Players?
Metric Huawei Nvidia
2025 Market Cap ~$160B (down from ~$500B in 2020, due to sanctions) ~$2.6
Trillion
Employees ~195,000 ~36,000
R&D Spending $24 billion $12.9 billion
HQ Shenzhen, China Santa Clara, USA
Notice how despite sanctions, Huawei invests almost 3x Nvidia?s R&D budget ? a
crucial factor in Ascend 910D?s rapid advancement.
Huawei Ascend 910D vs Nvidia H100: Where Huawei Stands
Globally, however, Huawei faces fierce rivals:
Nvidia H100 (5nm, TSMC):
Leading in performance, ecosystem (CUDA), and memory tech (HBM3).
Nvidia B100 (Expected 2025, 4nm, TSMC):
Projected to deliver 30?40% higher performance over H100, introducing HBM3e
memory.
AMD Instinct MI300X (6nm, TSMC):
Dominates memory bandwidth benchmarks, key for LLM training.
Cerebras and SambaNova:
Offering revolutionary wafer-scale and reconfigurable AI accelerators, although
still niche compared to Nvidia.
Thus, while the Ascend 910D is a giant leap for Huawei internally, the global
AI race remains crowded and intense.
Huawei?s strategy:
Focus first on China, Middle East, Russia, and countries less aligned with U.S.
trade policies.
Offer lower TCO (Total Cost of Ownership) solutions than Nvidia/AMD.
Push MindSpore as an alternative to CUDA, and pre-build LLMs fine-tuned for its
chip.
Detailed Technical Comparison: Why Ascend 910D Matters
Feature Ascend 910D Nvidia H100 AMD MI300X
Peak FP16 1.2 PFLOPS 1.0 PFLOPS 1.3 PFLOPS
Peak INT8 2.4 PFLOPS 2.0 PFLOPS 2.6 PFLOPS
Memory Bandwidth 800 GB/s (HBM2e) 3 TB/s (HBM3) 5.2 TB/s (HBM3)
Supported Model Size 175B parameters 530B parameters 500B parameters
Process Node 7nm (SMIC N+2) 4N (TSMC) 5nm (TSMC)
Energy Efficiency +12% better than H100 Baseline +10% over H100
TDP (Power Draw) 350W 700W 750W
? Strengths:
Lower power
Lower cost (~30?40% cheaper)
Sanctions-proof design
? Weaknesses:
Slower memory access (HBM2e vs HBM3)
Software ecosystem less mature
How Sanctions Shaped Huawei?s Strategy
Because SMIC (Huawei?s foundry partner) is restricted from buying advanced EUV
lithography machines, the Ascend 910D uses 7nm DUV manufacturing ? impressive
but behind Nvidia?s 4N process (based on TSMC?s 5nm EUV).
Yet Huawei:
Increased chip area to fit more transistors.
Focused on low-frequency operation to control heat.
Optimized matrix multiplication units for LLMs instead of broad AI workloads.
This is a design-for-purpose chip, not a general-purpose competitor yet.
Risks and Limitations for Huawei?s Ascend 910D
Despite its impressive debut, the Ascend 910D faces critical limitations:
Manufacturing Constraints:
Built at 7nm while rivals move to 5nm and even 4nm processes. Lower transistor
density affects long-term competitiveness.
Memory Bottlenecks:
Limited to older versions of HBM memory, affecting ultra-large model training.
Software Ecosystem:
MindSpore adoption is limited compared to Nvidia?s CUDA global dominance.
International Market Access:
U.S. sanctions restrict Huawei from selling freely outside China, Europe
remains cautious.
These factors could prevent Ascend 910D from replicating its domestic success
internationally.
Why Nvidia Still Dominates Globally
Nvidia remains the undisputed global AI chip leader, for several strategic
reasons:
Full Stack Advantage:
Beyond hardware, Nvidia controls the AI software world via CUDA, TensorRT, and
cuDNN libraries.
Strategic Cloud Alliances:
Deep integration with AWS, Microsoft Azure, Google Cloud ensures Nvidia
hardware is embedded in the world?s AI pipelines.
Innovation Pace:
Jensen Huang?s leadership pushes Nvidia to aggressive product refresh cycles ?
every 18?24 months.
Financial Power:
As of April 2025, Nvidia?s $2.2 trillion market cap dwarfs Huawei?s chip
division, which operates under strict budgetary and geopolitical constraints.
In simple terms, while Huawei is climbing, Nvidia built the mountain.
Still a long Battlefield
The launch of Ascend 910D underscores Huawei?s growing ability to innovate
under pressure.
It strengthens China?s domestic AI landscape and helps Huawei regain pride
after years of tech embargoes.
Yet globally, Nvidia?s lead remains decisive ? for now.
Over the next five years, the battle for AI chip supremacy will hinge not just
on designing better chips, but also on securing access to cutting-edge
manufacturing, building trusted ecosystems, and winning developer loyalty.
The real AI chip wars are just beginning.
------------------------------
Subject: Digest Footer
_______________________________________________
Link mailing list
[email protected]
https://mailman.anu.edu.au/mailman/listinfo/link
------------------------------
End of Link Digest, Vol 390, Issue 1
************************************
