Brian Atwood
@batwood011
Followers
23,815
Following
2,538
Media
320
Statuses
4,815
Creating conversational speech AI for the real world @sindarintech . Try for yourself at
Austin, TX
Joined June 2017
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#WWERaw
• 129274 Tweets
ブロック
• 127028 Tweets
Bengals
• 105672 Tweets
Jayden Daniels
• 79166 Tweets
うまい棒
• 75435 Tweets
連休明け
• 65674 Tweets
Terry
• 63474 Tweets
GFRIEND
• 54593 Tweets
呂布カルマ
• 50388 Tweets
Jey Uso
• 46462 Tweets
VOTA X KARIME
• 45424 Tweets
#沼におぼれる
• 20529 Tweets
ラッパー
• 19828 Tweets
lorde
• 19794 Tweets
イーロン
• 17466 Tweets
tini
• 16843 Tweets
インプレゾンビ
• 16638 Tweets
SATUIN SEMANGAT Pasti RIDO
• 16572 Tweets
SATUKAN Energi POSITIF Kita
• 15851 Tweets
クロススカウト
• 13758 Tweets
コレクション缶バッジ
• 11564 Tweets
I worked as a PM at a large tech co b/c after working really hard as a SWE I was mystified at what a PM actually does and wanted to learn
It was a 1.5 year highly paid vacation
Calling it now. Product Management as implemented over the last decade or so was a ZIRP.
81
95
2K
33
97
2K
Alright, the Devin hype is too much.
Devin is not an LLM.
Why are they comparing its performance against LLMs?
Of course GPT-4 in an agentic loop with full shell and browser access is going to beat a one-shot GPT-4 inference.
Please, someone hack together a gpt-4 dev loop
90
55
936
3:30am on a Sunday and we just cracked consistent sub-700ms voice-to-voice latency with ~perfect hearing.
Going into production in the coming months.
2024 is gonna be wild.
19
12
573
Slowly, but surely, everything we believed about the design and structure of software and software companies is melting away
16
34
492
Plot twist:
The Safety team left not because they saw *something* but because they saw *nothing*
No real danger. Only limitations, dead ends and endless distractions with commercialization — no path to AGI with GPT but the product is live and the bags must be pamped
AGI will
52
16
371
#coronavirus
stocks:
$GILD (has the cure, probably)
$CODX (has the test, probably)
$APT (protective gear mfctr)
$MRNA (vaccine)
$VIR (antibodies)
26
99
309
Python package management is truly a disaster to behold
38
12
260
There’s no conversational AI today that’s *enjoyable* to have a conversation with.
Not even close.
92
9
244
Within a year you’ll be able to have a spoken conversation with a computer that’s virtually indistinguishable from a conversation with a real person.
Within three years those conversations will feel better than most conversations with real humans.
27
19
179
It looks like OpenAI is gearing up to release Voice Assistants on Monday.
The moment we
@SindarinTech
have been anticipating for nearly two years.
Their quality will almost certainly beat everyone else’s, at least along specific axes.
But will they come to single-handedly to
18
10
184
@LucasFabisiak
@erdocAA
@peterboghossian
Women with abnormally large breasts don’t choose to put them on every morning before heading to work.
5
0
163
You should listen to Yacine.
Just ignore him when he’s shitposting
And don’t ignore him when he isn’t
Simple
20
3
177
12,000 people spoke with our conversational pitch deck this weekend
17
11
165
Turn-taking is one of the very hard problems in conversational speech AI.
A conversational agent that talks over you before you're done speaking immediately breaks immersion and discourages further engagement.
Here's a sneak preview of what we've been cooking.
Going in prod in
19
18
159
Just shaved another 150ms off end-to-end latency. In prod as of 10 minutes ago.
8
0
138
I’ve detested Humane for years.
Like many others here, some friends and I were tinkering with the “Personal AI” concept well before GPT-3 — wrestling with the Stanford NLP toolkit, writing parsers by hand, building SMS prototypes, etc — paying our dues.
When Humane’s vision,
5
5
127
Simply remembering to use an LLM is an S-tier engineering skill
6
5
122
@OfficialLoganK
All I want to know is: if a reasonably intelligent and capable team spent a few weeks building an agentic framework with the same tools (IDE, runtime, browser etc) and GPT-4, what score would they achieve on SWE-bench?
This chart doesn’t give us any insight into that question,
11
3
120
@AamerAnwar
@realDonaldTrump
There is nothing insane about this.
@realDonaldTrump
is completely, 100% right on this. WHO has done nothing but lie and misinform every step of the way on this.
7
2
86
It pains me to think of the number of idle H100s allocated for products nobody’s using
7
2
94
3
0
86
Quick primer on
@SindarinTech
:
We stay up late building the best possible realtime conversational speech engine for LLMs so that you don't have to.
Building async chatbots on LLMs is easy.
Building low-latency speech-based chatbots is *much* harder.
There's really one best
15
9
77
“Coding is dead” is the most low energy, unimaginative take
Coding is more fun and impactful now than ever before and it will only be more so thanks to LLMs
The ceiling for creative programmers is approaching outer space
coding isn't dying, just dying as we know it
once AI can build all the apps we have today easily, we'll simply start building more complex apps til they start breaking
coding work will look more like a car mechanic, you're not building the car or car parts, but you know where
128
232
2K
5
1
77
@mimi_aboubaker
@proetrie
Many ambitious men are already pushing themselves right up to their breaking point, even if it doesn’t look like it.
To the extent that a woman can tap greater potential, her best bet is to do it with great sensitivity, kindness, and patience.
5
0
72
@mitchellh
“This reality”?
JS *is* reality.
The Event Loop is the basis of consciousness.
Callbacks separate us from the animals.
String theorists only got two things wrong.
Its not strings.
It’s thread.
And it’s running on pid 11436.
3
3
71
I take it back multimodality is clearly the future
I don't know what GPT4o's voice-to-voice architecture is, but I can tell you everything we saw in today's demo can be done without an E2E model.
22
0
64
7
3
73
I feel bad for anyone who hasn’t bought
#Bitcoin
because they don’t see “intrinsic value”
A USD note doesn’t have “intrinsic value” beyond paper + ink
It’s about the properties of money, which
@NickSzabo4
has written extensively about
6
5
60
I don't know what GPT4o's voice-to-voice architecture is, but I can tell you everything we saw in today's demo can be done without an E2E model.
22
0
64
@gfodor
If you truly believe machines will reach AGI on short time horizon there isn’t much reason to do anything other than YOLO.
But LLMs (at least this crop) aren’t AGI and have very specific limitations that require enormous scaffolding to do anything actually useful aside from
5
4
65
A VC once said to me, perplexed:
“So you’re just allowing people to speak without holding a button down?”
Lol. Lmao, even
How do I detect when someone is done speaking when transcribing with something like Whisper?
cc
@batwood011
69
4
226
6
1
62
@pmarca
This isn’t a complete answer.
Pure wordcels operate in the scarce world of status dominance, requiring a zero-sum competitive disposition.
Shape rotators operate in the abundant world of shapes, where there are always plenty of shapes to go around.
5
3
55
@tunguz
Sure, some Python here and there, but Node.js will pull it all together
I don’t make the rules
5
1
57
@brianormou5
@yacineMTB
Serial -> parallel
Human generated -> AI generated
Deterministic -> nondeterministic
Static -> dynamic
All ways of pointing to the thing
4
2
54
Guess which AI voice engine *is* ready
So, basically what OpenAI demoed with Voice wasn't even **close** to being ready.
75
27
928
3
0
57
> can’t get SWE friend to try chatgpt
> SWE friend finally tries chatgpt
> kinda cool anon, but far from usable
> ask if they tried chatgpt4
> no, just the free version
> try 4 it’s so much better
> no reply, already made up their mind
> mfw
8
0
52
Large training runs should take great care to omit
@ESYudkowsky
’s precise humanity-annihilating instructions from their training sets.
Late night thought:
If LLMs or AI trained on our words are a true threat, and if the most powerful models will be trained on the latest data…
Then maybe spreading love and empathy is the best AI safety method. Every time we say something positive, we make it likely for future
124
189
1K
10
6
51
To whoever just deployed a script to ddos our pitch deck
Focus on yourself, king 👑
4
2
52
@primalpoly
@peterboghossian
@JonHaidt
As a psychologist you should know better.
The moral intuition here is identical to the intuition for why there should be dress codes in schools / workplaces at all.
It’s perfectly psychologically sound.
2
0
48
@balajis
Three weeks ago I was told remote court in Austin TX “couldn’t be done.”
Today it is 100% remote.
2
7
49
HOW TO CODE AT A STARTUP
An essay about how to think about coding at a startup.
PART I: Startup Philosophy
The most important thing to remember when working at a startup is the ratio of upside opportunity to downside risk.
That ratio should permeate every decision you make.
10
4
50
Traditional VC wisdom really needs revision when 10x engineers are made into 100x engineers by copilots
3
6
46
Hard to overstate how disruptive > GPT-4 models are going to be for society.
Take a walk or drive around any thriving / growing city.
Most of the wealth — the apartments, the offices, the stores, the restaurants, gyms, etc — depends on the incomes, and presumed income growth,
16
0
43
“Based” is short for “based in reality” — being based means first accepting that reality is real (a controversial claim!) then stating something about it that is probably true and useful yet uncouth in the mixed company of coastal elites
6
5
42
Wow, we just discovered a bug that was increasing response times by up to 500+ms for the last two months
Nobody even noticed because it’s already so fast
3
0
43
@batwood011
They were promised 20% of the GPUs to do their research but then they saw they needed to keep customers up and running and switched priorities.
Businesses need to serve customers first. Otherwise they go away.
3
2
11
5
2
41
Our conversational speech engine is now wicked fast.
Try for yourself on Chrome at
3
5
40
The most human thing you can do is try to get really rich off of AI just ahead of the singularity
3
2
40
@Austen
@awilkinson
Instead of researching bogus “vaccines” Gates could be giving every American $50 million.
1
0
37
Epic of Gilgamesh - 2100 BCE
Code of Hammurabi - 1754 BCE
Edicts of Ashoka - 232 BCE
Rosetta Stone - 196 BCE
Magna Carta - 1215 CE
Gutenberg Bible - 1454 CE
Declaration of Independence - 1776 CE
Llama 3 400b - 2024 CE
1
0
37
Due to increasing competition in the space, we've decided to pivot
@SindarinTech
In the interest of full transparency, I've attached a video of our most recent investor pitch
We're really excited about this new direction!
(Sound on 🔊)
3
3
36
@meaning_enjoyer
From this you can infer that most white-collar knowledge worker types are incessantly beating their meat to modulate their sexual desire to match the life-affirmingness of reading VLOOKUP docs for the 18th time
2
1
33
This video is incredible.
@Figure_robot
and
@OpenAI
did an amazing job end to end here.
Incidentally, I just spent 30 minutes with
@SindarinTech
to spin up a demo of the same with Sindarin Prime, our incorporeal company robot.
No OpenAI APIs involved.
With OpenAI, Figure 01 can now have full conversations with people
-OpenAI models provide high-level visual and language intelligence
-Figure neural networks deliver fast, low-level, dexterous robot actions
Everything in this video is a neural network:
1K
5K
17K
2
2
36
@proetrie
Do you think intelligent people struggle with it more than unintelligent people do?
1
0
30
I experienced something tonight that very few have experienced.
In a few weeks millions of others will experience it.
2
0
32
There were some really cool features in today’s demo!
And I don’t think we are dead.
8
1
33
@ItakGol
Who had “meta open sources LLM under commercial license, stability finetunes it and releases under non-commercial license” on their bingo card
4
0
31
It’s been less than 24 hours since we added an open source LLM to
@SindarinTech
and people are already creating AI girlfriends with it.
6
2
31
I’m hiring an ML engineer
@SindarinTech
to help with training/finetuning, optimizing, and deploying a number of models (mostly DL).
DM if interested
3
3
29
Hot take: E2E voice models like 4o are overrated and cascaded approaches are going to have much longer lifespans — and much greater capability — than almost anyone thinks.
And yes, I’ve read The Bitter Lesson.
13
2
30
Something you quickly notice when speaking with an embodied AI is how attentive they are.
Unlike humans, they don’t split attention. You are their first and only priority.
It feels incredible and sets an incredibly — if not impossibly — high bar for humans as conversational
5
3
29
@academic_exit
I literally believed in the efficient market hypothesis
Down to the micro operational level
Couldn’t believe that companies weren’t perfectly, platonically efficient
2
0
27