Kseniya Lukashuk

Building AI ecosystems, founder access & global partnerships @function1_conf @thehypedotnews

[thehype. @thehypedotnews]

gpt astra vs fable 5.1 at goldberg machine

gpt 6 astra – openai, landed on @OpenRouter less then hour ago, provider pinned to openai
fable 5.1 – anthropic, shipped sep 1

we put the two models on one job: a rube goldberg machine in three.js that presses a button and detonates a bomb

the setup: one self-contained html file, three.js from a cdn, everything else procedural – no textures, no models, no physics engine, every collision hand-written. the hard part sits in the brief: a domino may only fall once the previous one actually touches it, checked by real overlap every frame, never by a timer. same rule for the hammer hitting the button and the button firing the bomb. one continuous camera, its speed driven by whatever is moving. we recorded both scenes frame by frame – 1200 frames, 60 fps, exactly 20 seconds – and stepped both by hand to read the telemetry.

- cost
#1 astra – $1.84
#2 fable – $29.16

- time
#1 astra – 9m 56s
#2 fable – 1h 12m

- tokens
#1 astra – 45k
#2 fable – 360k

- lines of code
astra – 881
fable – 744

observations:

• we told it what we saw and nothing else – no diagnosis, no patch. we never edit a model's code. round two ran the whole chain to the blast.

• both files are deterministic. two runs each, identical state to twelve decimals, and neither model reached for math.random.

conclusion: 15.8x cheaper and 7.2x faster, and it still took a second round to get the ball into the bucket!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

@thehypedotnews @OpenRouter creepy animals tbh

[Kseniya Lukashuk @KsiuShines]

@bridgemindai fable 5.1 is crazy https://t.co/uJlRe9KDDu

[thehype. @thehypedotnews]

one prompt, 43 minutes, $2.96 – claude fable 5.1 wrote a 1,660-line world

we gave it one job: build a detailed frutiger aero world in a single self-contained html file, with a camera travelling through it for 600 frames

the setup: @threejs 0.169 and nothing else – no images, no downloaded textures, no physics library. every surface drawn in code. the file has to expose reset/step/simstate, render 600 frames and produce the same picture twice. one pass. every number below is ours, measured on the rendered frames

what came back:

• 1,660 lines in one file, 61,363 characters
• 60,000 instanced blades of grass, 500 flowers
• 400 towers, 700 canopy pieces, 90 rocks
• 60 bubbles in the system, 30 on screen at frame 600
• 1.29m triangles at 47 draw calls
• a 132m camera path across 8 control points
• rendered at 2x and downsampled

observations:

• it seeded its generator with 20050607 – june 7, 2005, the year the look it was copying belongs to
• 47 draw calls for 1.29m triangles. it instanced everything that repeats, without being told which
• the brightest frame is 3.6% pure white. the brief capped it at 25%, and this look usually blows out

conclusion: one file, no assets, 1.29m triangles and a camera that walks 132 metres through it, for $2.96!

follow @thehypedotnews for 24/7 ai news, analysis and breakdown

[Kseniya Lukashuk @KsiuShines]

@Bhavani_00007 fable 5.1 is soooo good at 3d building! https://t.co/uJlRe9KDDu

[thehype. @thehypedotnews]

one prompt, 43 minutes, $2.96 – claude fable 5.1 wrote a 1,660-line world

we gave it one job: build a detailed frutiger aero world in a single self-contained html file, with a camera travelling through it for 600 frames

the setup: @threejs 0.169 and nothing else – no images, no downloaded textures, no physics library. every surface drawn in code. the file has to expose reset/step/simstate, render 600 frames and produce the same picture twice. one pass. every number below is ours, measured on the rendered frames

what came back:

• 1,660 lines in one file, 61,363 characters
• 60,000 instanced blades of grass, 500 flowers
• 400 towers, 700 canopy pieces, 90 rocks
• 60 bubbles in the system, 30 on screen at frame 600
• 1.29m triangles at 47 draw calls
• a 132m camera path across 8 control points
• rendered at 2x and downsampled

observations:

• it seeded its generator with 20050607 – june 7, 2005, the year the look it was copying belongs to
• 47 draw calls for 1.29m triangles. it instanced everything that repeats, without being told which
• the brightest frame is 3.6% pure white. the brief capped it at 25%, and this look usually blows out

conclusion: one file, no assets, 1.29m triangles and a camera that walks 132 metres through it, for $2.96!

follow @thehypedotnews for 24/7 ai news, analysis and breakdown

[Bhavy☄️ @Bhavani_00007]

I tested Fable 5.1 vs Kimi K3

I gave 3 same prompts to both, here are the results

IMO, both created beautiful scenes, but the color picking problem is still there in Fable 5.1. Kimi K3 is better at this.

Fable: ~$12/scene | Kimi: <$5/scene

More tests coming up

[Kseniya Lukashuk @KsiuShines]

@claudeai testing https://t.co/uJlRe9KDDu

[thehype. @thehypedotnews]

one prompt, 43 minutes, $2.96 – claude fable 5.1 wrote a 1,660-line world

we gave it one job: build a detailed frutiger aero world in a single self-contained html file, with a camera travelling through it for 600 frames

the setup: @threejs 0.169 and nothing else – no images, no downloaded textures, no physics library. every surface drawn in code. the file has to expose reset/step/simstate, render 600 frames and produce the same picture twice. one pass. every number below is ours, measured on the rendered frames

what came back:

• 1,660 lines in one file, 61,363 characters
• 60,000 instanced blades of grass, 500 flowers
• 400 towers, 700 canopy pieces, 90 rocks
• 60 bubbles in the system, 30 on screen at frame 600
• 1.29m triangles at 47 draw calls
• a 132m camera path across 8 control points
• rendered at 2x and downsampled

observations:

• it seeded its generator with 20050607 – june 7, 2005, the year the look it was copying belongs to
• 47 draw calls for 1.29m triangles. it instanced everything that repeats, without being told which
• the brightest frame is 3.6% pure white. the brief capped it at 25%, and this look usually blows out

conclusion: one file, no assets, 1.29m triangles and a camera that walks 132 metres through it, for $2.96!

follow @thehypedotnews for 24/7 ai news, analysis and breakdown

[Kseniya Lukashuk @KsiuShines]

@thehypedotnews @threejs we r cooked

[Kseniya Lukashuk @KsiuShines]

fable 5.1 appears to be a nice texture builder

[thehype. @thehypedotnews]

one prompt, 43 minutes, $2.96 – claude fable 5.1 wrote a 1,660-line world

we gave it one job: build a detailed frutiger aero world in a single self-contained html file, with a camera travelling through it for 600 frames

the setup: @threejs 0.169 and nothing else – no images, no downloaded textures, no physics library. every surface drawn in code. the file has to expose reset/step/simstate, render 600 frames and produce the same picture twice. one pass. every number below is ours, measured on the rendered frames

what came back:

• 1,660 lines in one file, 61,363 characters
• 60,000 instanced blades of grass, 500 flowers
• 400 towers, 700 canopy pieces, 90 rocks
• 60 bubbles in the system, 30 on screen at frame 600
• 1.29m triangles at 47 draw calls
• a 132m camera path across 8 control points
• rendered at 2x and downsampled

observations:

• it seeded its generator with 20050607 – june 7, 2005, the year the look it was copying belongs to
• 47 draw calls for 1.29m triangles. it instanced everything that repeats, without being told which
• the brightest frame is 3.6% pure white. the brief capped it at 25%, and this look usually blows out

conclusion: one file, no assets, 1.29m triangles and a camera that walks 132 metres through it, for $2.96!

follow @thehypedotnews for 24/7 ai news, analysis and breakdown

[Kseniya Lukashuk @KsiuShines]

@RoundtableSpace we generated a 3d world as well, fable 5.1 is good at threejs building https://t.co/uJlRe9KDDu

[thehype. @thehypedotnews]

one prompt, 43 minutes, $2.96 – claude fable 5.1 wrote a 1,660-line world

we gave it one job: build a detailed frutiger aero world in a single self-contained html file, with a camera travelling through it for 600 frames

the setup: @threejs 0.169 and nothing else – no images, no downloaded textures, no physics library. every surface drawn in code. the file has to expose reset/step/simstate, render 600 frames and produce the same picture twice. one pass. every number below is ours, measured on the rendered frames

what came back:

• 1,660 lines in one file, 61,363 characters
• 60,000 instanced blades of grass, 500 flowers
• 400 towers, 700 canopy pieces, 90 rocks
• 60 bubbles in the system, 30 on screen at frame 600
• 1.29m triangles at 47 draw calls
• a 132m camera path across 8 control points
• rendered at 2x and downsampled

observations:

• it seeded its generator with 20050607 – june 7, 2005, the year the look it was copying belongs to
• 47 draw calls for 1.29m triangles. it instanced everything that repeats, without being told which
• the brightest frame is 3.6% pure white. the brief capped it at 25%, and this look usually blows out

conclusion: one file, no assets, 1.29m triangles and a camera that walks 132 metres through it, for $2.96!

follow @thehypedotnews for 24/7 ai news, analysis and breakdown

[thehype. @thehypedotnews]

hy4 preview vs glm 5.3 vs qwen 3.8 max

hy4 shipped all three for $0.537 – 3.2x under qwen. glm broke the fewest times, 2 fixes against hy4's 5. qwen wrote the most code, burned the most tokens, and still ships one of its three builds broken

the setup: three prompts of ~600 lines each, one html file per build, everything procedural – no meshes, no image files, no libraries past @threejs

prompts:

1. the petronas twin towers in dawn mist
2. taipei 101 in a tropical rainstorm
3. the cn tower in heavy snow

each file carries the geometry, the weather, night lighting, five scripted camera shots, a capture mode and a self-check panel that prints its own numbers. run through @OpenRouter, one shot per model, no reference images – two of the three models have no vision at all

models: @TencentHunyuan hy4 preview, @Alibaba_Qwen qwen 3.8 max, @Zai_org glm 5.3

- total cost, three builds
#1 hy4 preview – $0.537
#2 glm 5.3 – $1.111
#3 qwen 3.8 max – $1.737

- wall clock, three builds
#1 glm 5.3 – 44m
#2 hy4 preview – 50m
#3 qwen 3.8 max – 102m

- output tokens
#1 hy4 preview – 186,635
#2 glm 5.3 – 231,481
#3 qwen 3.8 max – 248,655

- fixes needed to make it run
#1 glm 5.3 – 2
#2 qwen 3.8 max – 4
#3 hy4 preview – 5

- lines of code shipped
#1 hy4 preview – 1,798
#2 glm 5.3 – 2,766
#3 qwen 3.8 max – 3,066

observations:

• every bug was one or two lines. no model failed the architecture – the geometry, the camera rigs and the self-check math were right everywhere. they broke on things a single run catches

• glm's petronas attempt spent 148,574 output tokens on reasoning and emitted zero characters of code. capping its thinking budget at 26k re-ran the same prompt in 683s for $0.250 – 3.2x faster and 2.7x cheaper

• no model won two towers in a row. taipei went to glm, petronas to qwen, cn tower back to glm, and the failures move the same way. the spread between tasks is bigger than the spread between models

conclusion: nine towers, 7,630 lines and 666,771 output tokens for $3.38 all in – and the cheapest model got there on 41% less code than the priciest!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[thehype. @thehypedotnews]

glm 5.3 flash is 7.5x cheaper, but 3.4x slower than gemini 3.7 flash

@Zai_org glm 5.3 flash – shipped aug 26, $0.07/$0.25 per 1m
@GoogleDeepMind gemini 3.7 flash – shipped aug 13, $0.38/$1.88 per 1m

we put the two models on one job: write one html file that draws an animated 3d scene in the browser. no images, no downloads, and it has to look the same on every load.

the setup: three scenes – a glass aquarium in a lit room, the solar system, a night city under a thunderstorm. identical brief word for word, reasoning effort high, 64k output cap.

the numbers below are not the whole run. they cover the three scenes we kept – the best one per task from each model, the ones in the video.

- total generation time for the three scenes
#1 gemini 3.7 flash – 10m 36s
#2 glm 5.3 flash – 36m 30s

- tokens spent on those three scenes
#1 glm 5.3 flash – 110k
#2 gemini 3.7 flash – 111k

- cost of those three scenes
#1 glm 5.3 flash – $0.027
#2 gemini 3.7 flash – $0.202

observations:

• glm's first 10 attempts: 7 blank pages. it kept inventing short random helpers and forgetting to define one of them. the fix was one line in the brief: use exactly one random helper, named rand(), and don't invent shorthands next to it. next 12 attempts: 11 alive, 0 crashes.

• glm spends 66% of its output on reasoning, gemini 57%. that is the whole speed gap.

• gemini's storm came back as a black rectangle in 4 of 6 runs. glm's best storm has a branching bolt, lit rain and wet asphalt – for $0.01.

conclusion: same three scenes, same token spend – glm 5.3 flash billed $0.027 and took 36m 30s, gemini 3.7 flash billed $0.202 and took 10m 36s. glm wins gemini on price and made the best storm of the whole run

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[thehype. @thehypedotnews]

ox alpha vs deepseek v4 flash vision vs grok 4.6 vs gemini 3.7 flash vs – on photo-to-3d

four vision models got one photograph each and had to rebuild the place inside it as a @threejs scene. twelve scenes, twelve first-try runs, zero console errors

the setup: one reference photo per scene, sent as an image on @openrouter. the prompt never says what is in the picture – no "motel", no "bar", no "gas station". the model has to read the photo and rebuild it: layout, materials, hour of the day, and whatever is around the corner that the frame does not show

tasks – three photographs of early-2000s america:

1. a motel at night, neon pylon lit, snow on the ground
2. an old new york tavern interior, tin ceiling, tiled floor
3. an abandoned service station in the california desert, midday sun

each scene ships as one self-contained html file, procedural geometry and canvas textures only, no downloads. three timed camera shots, and shot 1 has to reproduce the framing of the reference photo

models: @xai grok 4.6, @googledeepmind gemini 3.7 flash, @deepseek_ai deepseek v4 flash vision exp, and ox alpha – a stealth model on openrouter, free, no lab attached to it yet

results:

- wall clock, three scenes
#1 gemini 3.7 flash – 11m 12s
#2 deepseek v4 flash – 15m 20s
#3 grok 4.6 – 28m 11s
#4 ox alpha – 38m 54s

- output tokens
#1 gemini 3.7 flash – 77,396
#2 ox alpha – 87,613
#3 grok 4.6 – 105,687
#4 deepseek v4 flash – 127,884

- lines of code shipped
#1 ox alpha – 2,090
#2 deepseek v4 flash – 2,291
#3 grok 4.6 – 3,529
#4 gemini 3.7 flash – 3,989

- total price
#1 ox alpha – $0.000
#2 deepseek v4 flash – $0.091
#3 gemini 3.7 flash – $0.136
#4 grok 4.6 – $0.697

observations:

• grok is 7.7x the price of deepseek. it is the only model that read the light – low sun, real shadows on the station, a cold night on the motel

• gemini is the fastest and the least deliberate. 17,158 reasoning tokens against deepseek's 99,172, and it still shipped the most code – 3,989 lines

• deepseek thought hardest and rendered plainest. 99,172 reasoning tokens, 5.8x gemini's, spent on layout rather than on light. its motel is the second best in the set for $0.030

• ox alpha is free and reads a photo as well as anything here – it lifted "family units / kitchenettes" off the pylon and redrew it in canvas

conclusion: twelve scenes, four models, zero fixes, and the whole run cost $0.924!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

@elonmusk humans are being replaced by ox alpha for $0 https://t.co/2Z8Dqk1Ico

[thehype. @thehypedotnews]

ox alpha vs deepseek v4 flash vision vs grok 4.6 vs gemini 3.7 flash vs – on photo-to-3d

four vision models got one photograph each and had to rebuild the place inside it as a @threejs scene. twelve scenes, twelve first-try runs, zero console errors

the setup: one reference photo per scene, sent as an image on @openrouter. the prompt never says what is in the picture – no "motel", no "bar", no "gas station". the model has to read the photo and rebuild it: layout, materials, hour of the day, and whatever is around the corner that the frame does not show

tasks – three photographs of early-2000s america:

1. a motel at night, neon pylon lit, snow on the ground
2. an old new york tavern interior, tin ceiling, tiled floor
3. an abandoned service station in the california desert, midday sun

each scene ships as one self-contained html file, procedural geometry and canvas textures only, no downloads. three timed camera shots, and shot 1 has to reproduce the framing of the reference photo

models: @xai grok 4.6, @googledeepmind gemini 3.7 flash, @deepseek_ai deepseek v4 flash vision exp, and ox alpha – a stealth model on openrouter, free, no lab attached to it yet

results:

- wall clock, three scenes
#1 gemini 3.7 flash – 11m 12s
#2 deepseek v4 flash – 15m 20s
#3 grok 4.6 – 28m 11s
#4 ox alpha – 38m 54s

- output tokens
#1 gemini 3.7 flash – 77,396
#2 ox alpha – 87,613
#3 grok 4.6 – 105,687
#4 deepseek v4 flash – 127,884

- lines of code shipped
#1 ox alpha – 2,090
#2 deepseek v4 flash – 2,291
#3 grok 4.6 – 3,529
#4 gemini 3.7 flash – 3,989

- total price
#1 ox alpha – $0.000
#2 deepseek v4 flash – $0.091
#3 gemini 3.7 flash – $0.136
#4 grok 4.6 – $0.697

observations:

• grok is 7.7x the price of deepseek. it is the only model that read the light – low sun, real shadows on the station, a cold night on the motel

• gemini is the fastest and the least deliberate. 17,158 reasoning tokens against deepseek's 99,172, and it still shipped the most code – 3,989 lines

• deepseek thought hardest and rendered plainest. 99,172 reasoning tokens, 5.8x gemini's, spent on layout rather than on light. its motel is the second best in the set for $0.030

• ox alpha is free and reads a photo as well as anything here – it lifted "family units / kitchenettes" off the pylon and redrew it in canvas

conclusion: twelve scenes, four models, zero fixes, and the whole run cost $0.924!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

@OpenRouter ox alpha is the new free king https://t.co/2Z8Dqk1Ico

[thehype. @thehypedotnews]

ox alpha vs deepseek v4 flash vision vs grok 4.6 vs gemini 3.7 flash vs – on photo-to-3d

four vision models got one photograph each and had to rebuild the place inside it as a @threejs scene. twelve scenes, twelve first-try runs, zero console errors

the setup: one reference photo per scene, sent as an image on @openrouter. the prompt never says what is in the picture – no "motel", no "bar", no "gas station". the model has to read the photo and rebuild it: layout, materials, hour of the day, and whatever is around the corner that the frame does not show

tasks – three photographs of early-2000s america:

1. a motel at night, neon pylon lit, snow on the ground
2. an old new york tavern interior, tin ceiling, tiled floor
3. an abandoned service station in the california desert, midday sun

each scene ships as one self-contained html file, procedural geometry and canvas textures only, no downloads. three timed camera shots, and shot 1 has to reproduce the framing of the reference photo

models: @xai grok 4.6, @googledeepmind gemini 3.7 flash, @deepseek_ai deepseek v4 flash vision exp, and ox alpha – a stealth model on openrouter, free, no lab attached to it yet

results:

- wall clock, three scenes
#1 gemini 3.7 flash – 11m 12s
#2 deepseek v4 flash – 15m 20s
#3 grok 4.6 – 28m 11s
#4 ox alpha – 38m 54s

- output tokens
#1 gemini 3.7 flash – 77,396
#2 ox alpha – 87,613
#3 grok 4.6 – 105,687
#4 deepseek v4 flash – 127,884

- lines of code shipped
#1 ox alpha – 2,090
#2 deepseek v4 flash – 2,291
#3 grok 4.6 – 3,529
#4 gemini 3.7 flash – 3,989

- total price
#1 ox alpha – $0.000
#2 deepseek v4 flash – $0.091
#3 gemini 3.7 flash – $0.136
#4 grok 4.6 – $0.697

observations:

• grok is 7.7x the price of deepseek. it is the only model that read the light – low sun, real shadows on the station, a cold night on the motel

• gemini is the fastest and the least deliberate. 17,158 reasoning tokens against deepseek's 99,172, and it still shipped the most code – 3,989 lines

• deepseek thought hardest and rendered plainest. 99,172 reasoning tokens, 5.8x gemini's, spent on layout rather than on light. its motel is the second best in the set for $0.030

• ox alpha is free and reads a photo as well as anything here – it lifted "family units / kitchenettes" off the pylon and redrew it in canvas

conclusion: twelve scenes, four models, zero fixes, and the whole run cost $0.924!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

new benchmark!

[thehype. @thehypedotnews]

ox alpha vs deepseek v4 flash vision vs grok 4.6 vs gemini 3.7 flash vs – on photo-to-3d

four vision models got one photograph each and had to rebuild the place inside it as a @threejs scene. twelve scenes, twelve first-try runs, zero console errors

the setup: one reference photo per scene, sent as an image on @openrouter. the prompt never says what is in the picture – no "motel", no "bar", no "gas station". the model has to read the photo and rebuild it: layout, materials, hour of the day, and whatever is around the corner that the frame does not show

tasks – three photographs of early-2000s america:

1. a motel at night, neon pylon lit, snow on the ground
2. an old new york tavern interior, tin ceiling, tiled floor
3. an abandoned service station in the california desert, midday sun

each scene ships as one self-contained html file, procedural geometry and canvas textures only, no downloads. three timed camera shots, and shot 1 has to reproduce the framing of the reference photo

models: @xai grok 4.6, @googledeepmind gemini 3.7 flash, @deepseek_ai deepseek v4 flash vision exp, and ox alpha – a stealth model on openrouter, free, no lab attached to it yet

results:

- wall clock, three scenes
#1 gemini 3.7 flash – 11m 12s
#2 deepseek v4 flash – 15m 20s
#3 grok 4.6 – 28m 11s
#4 ox alpha – 38m 54s

- output tokens
#1 gemini 3.7 flash – 77,396
#2 ox alpha – 87,613
#3 grok 4.6 – 105,687
#4 deepseek v4 flash – 127,884

- lines of code shipped
#1 ox alpha – 2,090
#2 deepseek v4 flash – 2,291
#3 grok 4.6 – 3,529
#4 gemini 3.7 flash – 3,989

- total price
#1 ox alpha – $0.000
#2 deepseek v4 flash – $0.091
#3 gemini 3.7 flash – $0.136
#4 grok 4.6 – $0.697

observations:

• grok is 7.7x the price of deepseek. it is the only model that read the light – low sun, real shadows on the station, a cold night on the motel

• gemini is the fastest and the least deliberate. 17,158 reasoning tokens against deepseek's 99,172, and it still shipped the most code – 3,989 lines

• deepseek thought hardest and rendered plainest. 99,172 reasoning tokens, 5.8x gemini's, spent on layout rather than on light. its motel is the second best in the set for $0.030

• ox alpha is free and reads a photo as well as anything here – it lifted "family units / kitchenettes" off the pylon and redrew it in canvas

conclusion: twelve scenes, four models, zero fixes, and the whole run cost $0.924!

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

@thehypedotnews @threejs 🇺🇸🦅

[thehype. @thehypedotnews]

glm 5.3 vs kimi k3 – four hogwarts locations in @threejs

the setup: streamed straight off @openrouter, no agent loop, no tools, no filesystem – one prompt in, one html file out. the system prompt is a three.js skill, 32,014 chars, byte-identical for both

models: @zai_org glm 5.3, @kimi_moonshot kimi k3

scenes: the castle from the black lake at night, the great hall by candlelight, the grand staircase, the quidditch pitch at midday. four camera shots each. glm crashed on the staircase, kimi crashed on the castle, one hand-fix each

- total build time, four scenes
#1 glm 5.3 – 81m 13s
#2 kimi k3 – 86m 10s

- total output tokens
#1 kimi k3 – 211,759
#2 glm 5.3 – 364,565

- total price
#1 glm 5.3 – $1.790
#2 kimi k3 – $3.259

- lines shipped
#1 glm 5.3 – 4,257
#2 kimi k3 – 4,309

observations:

• kimi is 3.1x more expensive per output token across all four scenes, with no outlier. it is the price, not one bad run

• kimi's speed is a serving question. 89.7 tok/s through @togethercompute, 31.5 through moonshot ai, same weights, 2.9x apart

• glm goes deeper on detail when it has room: 1,463 lines on the castle, real stone coursing, window reveals with mullions, dormers cut into the slate. it is also the one that plans hardest – 250-320k chars of reasoning before it writes a line

follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

[Kseniya Lukashuk @KsiuShines]

@NousResearch 🔥

[Kseniya Lukashuk @KsiuShines]

@humynlabs loooved the video 😄...and this is why real-world data matters:)

[Humyn Labs @humynlabs]

We're Humyn Labs. With us, every robot works fine.

We turn human experience into robot skills built from thousands of hours of model-ready data.

Unlike LLMs, that learnt from the entire internet, robotics data doesn't get that shortcut.

Robots need real-world experience to learn how to:
- Iron clothes with a toddler underfoot and a dog at your heels
- Restock shelves while navigating a crowded, constantly changing store
- Assemble components amid the variability of a real factory floor

That experience happens once, in the physical world, and disappears.

We’re building the data platform Physical AI learns from: four sensory modalities, real-world environments, and the human experience needed to turn moments into skills.