Kind: captions
Language: en
You're probably checking the same AI
news sites every morning wondering if
you've missed something big. Well, I
spent the last week diving deep into
every major AI announcement. And here's
what surprised me. The most important
breakthroughs aren't coming from where
you'd expect. One of them is completely
free. And it's already outperforming
systems that cost millions to access.
Welcome back to bitbiased.ai,
where we do the research so you don't
have to. Join our community of AI
enthusiasts with our free weekly
newsletter. Click the link in the
description below to subscribe. You will
get the key AI news, tools, and learning
resources to stay ahead.
So, in this video, I'm breaking down
seven gamechanging AI updates from the
past week that you actually need to know
about. From an open-source model that's
beating MIT's toughest math problems to
AI glasses you can actually afford,
we're covering what matters and why it
matters to you. First up, let's talk
about the open- source breakthrough
that's making everyone rethink the
future of AI development. Story one,
Deepseek Math crushes the competition.
Here's something that should make you
pay attention. Deepseek just dropped a
math-solving AI that scored 118 out of
120 on the Putinham exam. Now, if you're
not familiar with the Putinham, think of
it as the Olympics of mathematics for
undergrads. It's brutally hard. We're
talking about problems that can stump
PhD candidates. But here's where it gets
interesting. This isn't some lockedway
proprietary system you need special
access to use. Deepseek Math V2 is
completely open- source, sitting right
there on GitHub for anyone to download
and experiment with. And it's not just
good at textbook problems. This thing
solved five out of six questions from
the International Mathematical Olympiad
2025, putting it in the same league as
the world's top human mathematicians.
What makes this different from your
typical language model is the
architecture itself. Instead of just
generating an answer and hoping for the
best, DeepSeek uses what they call a
generator verifier framework. One part
of the model attempts the solution while
another part checks the work and catches
mistakes.
Think of it like having a math genius
with a built-in fact checker running in
real time, constantly debugging its own
reasoning.
Now, this matters for a bigger reason
than just solving hard math problems.
While companies like OpenAI and Google
are keeping their most advanced models
behind closed doors, DeepSeek is proving
you can match or beat proprietary
systems and still make everything
public.
Researchers are already predicting this
could accelerate breakthroughs across
physics, engineering, and scientific
research because suddenly anyone with
decent hardware can access frontier
level mathematical reasoning. The models
available right now on GitHub and major
ML platforms.
Whether you're a student trying to
understand complex proofs, an engineer
running simulations, or a researcher
testing hypothesis, you've got access to
one of the strongest reasoning engines
ever built.
And that shift from closed to open,
that's going to reshape who gets to
participate in the next wave of AI
innovation.
Story two, Perplexity gets smarter about
your life. Perplexity just rolled out
two features that push it way beyond
being just another search engine. And
honestly, one of them changes how you'll
think about using AI for daily work.
First, the practical one, multicar
support. If you're like most people
juggling work and personal schedules
across Gmail, Outlook, maybe a third
account for side projects, you know the
headache of constantly switching between
apps. Perplexity's assistant now handles
all of them simultaneously.
It can schedule meetings, draft email
responses, spot scheduling conflicts,
and coordinate time blocks across every
calendar you connect.
No more double booking yourself because
you forgot to check your other inbox.
But here's the update that really stands
out. Persistent memory. This isn't just
the AI remembering your last few
questions.
Perplexity can now remember your
preferences, your writing style, the
products you care about, the research
topics you keep coming back to, even the
way you like information formatted. Once
you tell it you prefer concise emails,
or that you're focused on specific tech
categories, it keeps that context and
applies it automatically going forward.
Wait until you see how this actually
works in practice.
Instead of reexplaining your needs every
time you start a new conversation, the
AI already knows. It understands your
tone. It remembers which tasks you do
regularly. And before you ask, yes, you
control this completely. You can view
everything it's stored, edit details, or
wipe the memory clean whenever you want.
This next part will surprise you.
Analysts are saying this memory system
puts perplexity in direct competition
with OpenAI's assistance, Google's
Gemini projects and Anthropic's clawed
workflows.
That's a pretty bold claim for what
started as a search tool. But when you
combine calendar management, email
handling, and an AI that actually learns
your preferences over time, you're
looking at something that feels less
like a chatbot and more like a personal
assistant who's been working with you
for months. The updates are rolling out
now to prous users with broader
availability coming soon. If you've been
looking for an AI that doesn't treat
every conversation like meeting you for
the first time, this is worth checking
out. Story three. Karpathy says,
"Schools are fighting the wrong battle."
Former Open AI researcher Andre Karpathy
just dropped a take that's making a lot
of educators uncomfortable. His message,
"Stop trying to detect AI generated
homework. It's not going to work and
schools are wasting time and resources
fighting a battle they've already lost.
Here's the core argument. Detection
tools, the ones schools are currently
spending money on, are fundamentally
broken.
They produce false positives. They miss
obvious AI content. And as models get
better, these tools become increasingly
useless.
Carpathy compared it to a technological
arms race that schools can't win. For
every detector that gets better, AI
models get better at evading them. But
here's where it gets interesting.
He pointed to something specific.
Google's upcoming models that can not
only solve exam questions perfectly, but
can mimic a student's handwriting style.
Think about that for a second.
If an AI can write an answer in your
exact handwriting, making it look like
you solved it on paper, how do you
detect that? You can't. And that's
exactly Karpathy's point. The real
danger, according to him, isn't that
some students will cheat. It's that
schools will start falsely accusing
students of using AI when they didn't.
As detectors get less reliable,
authentic student work will trigger
flags, leading to wrongful accusations,
appeals, and potentially serious
consequences for students who did
nothing wrong. So, what's his solution?
Move graded assessments back into the
classroom. Let teachers directly observe
how students think through problems.
Keep homework as practice where AI tools
can be used openly as learning aids, not
banned outright.
Embrace AI as a study companion, but
make sure students can also function
without it when it counts.
This isn't just philosophical talk.
Schools worldwide are already dealing
with lawsuits over false AI accusations.
Detection companies are facing criticism
for unreliable tools. Teachers are
burned out from trying to police every
assignment.
Carpathy's stance is controversial, but
a growing number of educators are
starting to ask the same question. What
if we've been approaching this entire
problem the wrong way? If you're a
student, teacher, or parent, this debate
is worth following. The shift Karpathy
is proposing could completely change how
homework and testing work in the next
few years. Story four. Alibaba wants AI
glasses in your daily life. Alibaba just
entered the AI wearables market and
they're not aiming for the high-end
luxury crowd. Their new Quark AI glasses
start at about $268,
making them one of the most affordable
smart eyeear options with serious AI
capabilities packed in. These aren't
just notification glasses.
They've got a built-in camera,
microphone array, and a lightweight
display. All powered by Alibaba's Quen
language models and Quark Assistant.
You can use them for realtime
translation when traveling. Get scene
descriptions of what you're looking at,
identify landmarks, navigate hands-free,
or pull up search results without
touching your phone. What makes this
different from earlier smart glasses
attempts?
The AI reasoning is significantly
better.
Early reviewers in China are comparing
them to a blend of Meta's Rayban AI
glasses and the original Google Glass,
but with more natural language responses
and stronger contextual understanding.
You can ask the glasses to summarize
text you're reading, take notes during
meetings, answer questions based on your
environment, or draft emails, all
without pulling out a device. Here's the
part that could shake up the market. At
this price point, Alibaba's positioning
these as everyday assistants for
travelers, students, and professionals.
Not as a niche tech experiment, but as
something you'd actually wear daily. And
considering how fast wearable tech is
evolving, being first to market with
affordable, functional AI glasses could
be a major advantage. This puts Alibaba
in direct competition with Meta, Apple's
vision products, and a wave of startups
building AR devices.
The companies testing initial demand
domestically before expanding
internationally. But if adoption rates
are strong in China, expect these to hit
global markets fast. Whether AI glasses
become mainstream or remain a niche
product depends a lot on whether
companies can make them useful enough to
justify wearing everyday. Alibaba's
betting that $268 and genuinely helpful
AI features are the combination that
tips the scale. Beyond headlines. All
right, let's hit three quick updates
that flew under the radar but still
matter. Story five. Open AAI's analytics
partner gets breached. Open AAI
confirmed that Mix Panel, one of their
analytics providers, suffered a security
breach. Some API users had their names,
emails, locations, and device
information exposed. The good news, no
API keys, no billing details, no actual
project content got compromised. The bad
news, if your info was leaked, you're
now a target for highly specific fishing
attempts.
Open AAI's already pulled Mix Panel from
their systems and started an internal
security review. They're notifying
affected developers directly. Security
experts are pointing out that while this
breach itself is limited, it highlights
a growing risk. Third-party tools
integrated into AI platforms are
becoming attack vectors.
As AI companies handle more sensitive
enterprise data, every connected service
becomes a potential weak point. If
you're an OpenAI API user and you get an
email about this, don't ignore it.
Update your security practices. Watch
for suspicious emails and enable
two-factor authentication if you haven't
already. Story six. Wait, didn't we just
cover this? You're not going crazy.
Story six in the original notes is
identical to story five. Looks like a
copype error in the source material.
Moving right along.
Story 7. Nvidia's tool orchestra proves
smaller can be smarter. Here's something
that challenges the bigger is always
better mentality dominating AI
development. NVIDIA and the University
of Hong Kong built a system called Tool
Orchestra that trains smaller models to
be strategic about when to think on
their own and when to call external
tools for help. The results are wild. An
8 billion parameter model using Tool
Orchestra scored 37.1%
on humanity's last exam, a benchmark
specifically designed to be extremely
difficult and beat GPT5 and Claude Opus
4.1 in the process.
And it did this while being two and a
half times more efficient and faster
than the larger models. But here's where
it gets interesting. Even when given
tools it had never seen before, the
model adapted seamlessly.
That suggests we might be entering a
phase where intelligence isn't just
about model size. It's about
coordination, tool selection, and
knowing when to outsource tasks versus
solving them internally. If this
approach scales, it could completely
shift how we build AI systems.
Instead of training bigger and bigger
models that try to do everything, we
might see ecosystems of smaller
specialized models that know how to work
together intelligently.
That would make AI development cheaper,
faster, and more accessible to teams
that can't afford to train frontier
models.
So, that's seven updates that actually
matter from this past week. From open-
source math models beating the toughest
exams to AI glasses you can afford,
we're watching intelligence get more
powerful, more accessible, and more
integrated into everyday tools.
If any of these caught your attention,
drop a comment. Are you excited about
open source models closing the gap with
proprietary systems? Do you think AI
glasses will actually become mainstream,
or are we still a few years out? And if
you're in education, how should schools
actually handle the AI homework
situation?
I'll see you in the next one.